Improving the Quality of Motorcycle Travel Data Collection (2013)

Below is the uncorrected machine-read text of this chapter, intended to provide our own search engines and external engines with highly rich, chapter-representative searchable text of each book. Because it is UNCORRECTED material, please consider the following text as a useful but insufficient proxy for the authoritative book pages.

15 Literature and Internet Review The organization of the literature and Internet findings presented in this chapter begins with existing technologies for motorcycle detection. For each technology, an explana- tion is given regarding how it is used and its advantages and disadvantages in relation to motorcycle data collec- tion. Findings from research projects that test and evaluate applications for improving motorcycle detection accuracy are discussed within the applicable technology section. New and/or promising technologies for improving the accuracy of motorcycle detection and classification follow. The final section in this chapter summarizes methodologies for esti- mating motorcycle VMT when motorcycle count data are unavailable. Sensors used for vehicle detection are generally classified as intrusive (in-roadway sensors) and non-intrusive (beside and/or over-roadway sensors). Intrusive sensors are installed directly on the pavement surface, in saw-cuts or bores in the road surface, by tunneling under the surface or by adher- ing directly to the pavement surface. Examples of intrusive detectors are pneumatic road tubes, piezoelectric sensors, inductive loop detectors, and magnetometers. Non-intrusive sensors are mounted above the lane of traffic they are moni- toring or on the side of a roadway where they can monitor multiple lanes of traffic. Examples of non-intrusive sensors are video image processors, microwave radar, laser radar, pas- sive infrared, passive acoustic array, and combinations of sen- sor technologies such as passive IR and Doppler microwave sensors (17). Several of these technologies are used by state DOTs to col- lect motorcycle travel data for HPMS reporting. As Table 4 shows, most states reported using pneumatic road tubes to conduct short counts for motorcycles and piezoelectric sensors to conduct continuous counts. Several states also reported testing several technologies for motorcycle counts and data collection (18). Detection and Classification Challenges Achieving a successful motorcycle travel monitoring pro- gram involves two primary objectives: (1) finding and imple- menting technologies that can accurately detect motorcycles, and (2) determining where to count motorcycles and how often. These two objectives might be thought of as defining the basic technology and methodology of detection and clas- sification. Some of the challenging detection and classification issues that arise in a motorcycle travel monitoring program are (19): • Motorcycle definition • Spatial and temporal factors • Lane discipline • Vehicle size • Vehicle occlusion Motorcycle Definition. FHWA defines motorcycles in two categories: (1) larger motorcycles with two or three wheels, and (2) motorized bicycles, which include mopeds and scoot- ers that require registration. Some states have adopted FHWA’s definition of motorcycles, but others define motorcycles in other ways. Some states define them as vehicles with two or three wheels in contact with the ground, a seat or saddle for passengers with a sidecar or trailer, a steering handlebar, and no enclosure for the operator. Several states also have identi- fied criteria to differentiate motorcycles from mopeds based on vehicle speeds, engine displacement or horsepower, or wheel diameter. Given the discrepancies in definition, FHWA proposed to provide additional guidelines on the definition of motorcycles from mopeds and scooters (4). The April 28, 2011 issue of the Federal Register contains a Notice announc- ing the revision to FHWA’s guidance regarding state reporting of motorcycle registration information (20). Spatial and Temporal Factors. Accurate travel monitor- ing of motorcycles requires knowing where and when they C H A P T E R 3 Findings and Applications

16 spacing of only 73.5 in.; and they are only 2 ft to 3 ft shorter than those of several more conventional subcompacts. It is easier to distinguish motorcycles from subcompact autos on the basis of magnetic length (as measured by induc- tive loops) than on the basis of axle spacing. This is partly because the vehicles’ differences in physical length are greater than the differences in axle spacing, and partly because—for vehicles with low metal content—magnetic length is shorter than physical length. Motorcycles generally have a magnetic length that is at least 3 ft shorter than their physical length (18). Vehicle Occlusion. For roadside detection systems, large vehicles in a lane closer to the detector may prevent detection of smaller vehicles in the adjacent lane(s). Wheel occlusion is a similar concept, but it applies specifically to occlusion of vehi- cles in closer lanes for detection systems designed to detect tire/ wheel systems. Intrusive Detection Technologies Many states use intrusive detection technologies to obtain motorcycle counts for continuous and short duration count- ing and HPMS reporting requirements. Intrusive detectors include: • Pneumatic road tubes • Inductive loops • Piezoelectric sensors • Magnetic sensors travel. The locations and times most appropriate for counting motorcycles are not necessarily the same as those for other vehicle types. The counting methods currently used by most states are only partial solutions and reflect a set of assump- tions about motorcycle ridership that may not be stable and valid for all situations. One goal of NCHRP Project 08-81 was to identify a method, or methods, to better select sites and times for monitoring motorcycle travel to better reflect the spatial and temporal behavior of motorcyclists during the week and on weekends. Lane Discipline. Because motorcycles only cover part of a traffic lane, piezoelectric sensors, inductive loops, and other detectors that also cover only part of a lane might not detect them. In some areas, motorcyclists may operate between lanes or even on paved shoulders for short distances when the road is congested, also avoiding detection. In order to maximize the probability that the wheels of a Class 1 vehicle will be detected by a piezoelectric sensor, the sensor should extend, as nearly as possible, across the entire lane (18). This means replacing the 6 ft sensors now commonly used for traffic counting with full lane-width sensors. Vehicle Size. Existing sensors often cannot reliably dis- tinguish motorcycles from subcompact automobiles on the basis of axle spacing. The axle spacings of current Harley- Davidson motorcycles, for example, are between 63 in. and 66 in. These spacings are only a few inches shorter than those of the recently introduced Smart ForTwo car, which has an axle Number of States Reporting (n=24) Short Counts Continuous Counts Technology Tested Used Tested Used Intrusive Road tubes 13 20 - - Piezoelectric cable 3 4 9 17 Conventional inductive loops 6 2 4 8 Piezoelectric film 1 0 4 3 Inductive loop signatures 1 0 2 1 Quadrupole loops 1 0 1 0 Magnetometers 1 0 2 0 Non-Intrusive Manual 0 1 - - Radar 7 3 4 5 Video 1 2 2 1 Infrared (IR, including TIRTLs) 5 0 4 3 Acoustic 1 0 2 0 Source: Reference (18, p. 3-2). Table 4. Data collection technology used for motorcycle travel.

17 The electronics unit transmits energy to the wire loops, and the system behaves as a tuned electrical circuit with the loop wire and lead-in serving as inductive elements. When a vehicle passes over or stops within the wire loop, the conduc- tive metal induces eddy currents in the wire, which reduces the loop inductance (23). The reduction is measurable within the electronics unit and signals a detection. Inductive loop detectors provide vehicle passage, presence, count, speed, and occupancy data, and newer versions can provide vehicle classification based on specific metal detected within the vehicle (17). Many states use an additional sensing component with inductive loops for classification purposes and for detect- ing smaller vehicles like motorcycles—piezoelectric sensors (sometimes called piezos). Installers must cut a slot in the pavement, typically at a 90 degree angle to the direction of traffic and covering the full lane-width. Application of a force like a vehicle tire crossing the sensor generates an electrical charge that is proportional to the pressure exerted on the sensor. Piezoelectric sensors provide vehicle counts, detect vehicle weight and speed, and classify vehicles based on axle count and axle spacing (17). One caveat for using piezoelec- tric sensors for weight is that Class 1 piezoelectric sensors are required for generating weight data, whereas the cheaper and less accurate Class 2 sensors are typically used for classi- fication. The installing agency must decide before purchasing whether the sensors will be used for collecting weight data or classification data. Many jurisdictions combine inductive loop detectors with piezoelectric sensors in varying combinations for con- tinuous and short classification counts. The most common configurations are the LPL or PLP (piezo-loop-piezo) con- figurations. Using a pair of loops without piezos facilitates continuous classification of vehicles by length. The Virginia DOT adopted standards for loops and piezoelectric sensors to improve the detection of motorcycles. The Virginia DOT specification calls for installing loops with four turns of wire and no splices, and two piezos stacked one above the other in a single sawcut to cover the entire lane-width (18). Certain characteristics of motorcycles and actions of motorcyclists make accurate detection and classification dif- ficult using these technologies. Motorcycles are smaller and lighter and contain less metal than other vehicles, which results in undercounts of motorcycles. Motorcycles traveling in groups and especially in certain staggered patterns are par- ticularly challenging for inductive loop/piezoelectric sensor detection systems (24). Motorcyclists often avoid riding in the middle part of the lane because of debris, oil, coolant, and other slick fluids. Detectors that cover only part of a lane might not detect motor- cycles that pass over the part of the lane that is not covered. These sensors represent applications of mature technolo- gies to traffic surveillance, but all have limitations when they are used to count and classify motorcycles. The major draw- backs of using intrusive devices are traffic disruptions for installation and repair, system failures associated with instal- lations in poor road surfaces, and use of substandard instal- lation procedures (21). Resurfacing of roadways and utility repair also can damage sensors and force the operating agency to reinstall these types of sensors (22). Pneumatic Road Tubes. Pneumatic road tubes are anchored directly to the pavement surface. A vehicle passing over the tube causes a burst of air pressure to travel through the rubber tube. Road tubes detect volume and speed and classify vehicles by axle count and spacing. They differ from the other three types of detectors in that pneumatic road tubes typically are used for short-term traffic counts of 1 or 2 days on roads with low to moderate traffic volumes (17). Even so, pneumatic tubes require close surveillance to ensure proper performance. Road tubes are one of the main types of sensors used for conducting short-term counts of traffic, including motor- cycles. Certain limitations of this technology can lead to both undercounting and overcounting of motorcycles. In particu- lar, road tube systems may undercount motorcycles because they have trouble distinguishing groups of motorcycles and detecting very lightweight motorcycles or because motor- cyclists may choose to steer around the tube, avoiding detec- tion altogether. On the other hand, because road tube systems have difficulty differentiating between subcompact vehicles and larger motorcycles, many subcompact vehicles (e.g., the Smart ForTwo and the Mini Cooper) may be incorrectly clas- sified as motorcycles. When road tubes are used with an 8 ft threshold (as is the current practice in at least some states), such misclassifications may result in appreciable overcount- ing of the number of motorcycles (17). Advantages of pneumatic road tubes include that they are quick to install for temporary recording of data and they are relatively inexpensive and easy to maintain. Negative attri- butes of road tubes include limited lane coverage and the fact that their efficiency is subject to weather, temperature, and traffic conditions (17, 21). Inductive Loops/Piezoelectric Sensors. Inductive loop detectors consist of the following primary components: • One or more turns of insulated wire placed in a shallow slot sawed in the pavement • An electronics unit located in a nearby cabinet or weather- proof housing • Lead-in cable from the edge of the roadway to the roadside electronics

18 There are two types of magnetic field sensors: (1) dual-axis and three-axis magnetometers, which detect changes in the vertical and horizontal components of the earth’s magnetic field produced by a ferrous metal vehicle and (2) induction magnetometers, sometimes called magnetic detectors, which measure changes in the magnetic flux lines when metal com- ponents in a vehicle travel past the detection zone (26). Advantages of dual-axis or three-axis magnetometers include that they are less susceptible than are induction loops to the stresses of traffic and pavement flexing so they are use- ful in places where loops are not feasible (e.g., bridge decks). Some induction magnetometers transmit data over a wireless radio frequency link. Some models also can be installed under the roadway without the need for pavement cuts. The down side is that induction magnetometers cannot detect stopped vehicles, and some models have small detection zones (17). Recent research on magnetic detector technology has resulted in the development of a wireless magnetic sensor network based on magneto-resistive sensor technology that can detect vehicles, including motorcycles (27). The develop- ers claim that the sensor is able to achieve high accuracy for motorcycle detection because the magnetic length of motor- cycles is clearly distinguishable from other vehicle types. Non-Intrusive Detection Technologies Non-intrusive detection technologies are sensors mounted to the side of the roadway, above the roadway, or installed beneath the pavement. Non-intrusive technologies cause mini- mal disruption to normal traffic operations during installation, operation, and maintenance compared to conventional (intrusive) detection methods. Some viable non-intrusive technologies for motorcycle detection are: • Video-based detectors • Microwave radar detectors • Laser radar detectors • Passive IR detectors • Passive acoustic detectors • Combinations of sensor technologies In general, these sensors measure vehicle counts, pres- ence, and passage. Some sensors also provide vehicle speed, vehicle classification, and multiple-lane, multiple-detection zone coverage (17). Video-Based Detectors. Video-based detectors use a computer to analyze the image input from a video camera based on different approaches. Some detectors analyze the video image of a vehicle that passes through a target area and determines detection based on the change in pixels within the detection zone. This type of system is sometimes called Some motorcyclists even operate between lanes (“lane split- ting”) or on shoulders and similarly avoid detection. The use of wide loops and full lane-width piezos (or wide loops alone) helps reduce errors in counts of motorcycles (18). However, the difficulty in detecting motorcycles that travel side-by-side and/or between lanes or on shoulders likely will continue to result in undercounts. Some advantages of using piezoelectric sensors with loops over using inductive loops alone are improved speed accu- racy, the ability to determine the classification of the vehicle based on axle spacing, and the ability to determine and moni- tor the weights of vehicles when used with weigh-in-motion (WIM) electronics (17). The down side is that piezoelectric sensors are subject to failure with little or no warning, espe- cially if installed in poor pavement. The detection accuracy of light vehicles such as motorcycles is reduced as these sensors age (17). Piezos, loops, and other detectors that cover only part of a lane may fail to detect motorcycles and have dif- ficulty detecting motorcycles in groups. Advantages of using inductive loops alone include: • Flexible design (shape and number of wire turns) to satisfy a large variety of applications • Well-understood technology with a large experience base • Insensitive to inclement weather such as rain, fog, and snow • High accuracy for count data as compared with other com- monly used techniques Research is underway at the University of Oklahoma to test and evaluate a microprocessor-based system that uses a sen- sor consisting of a single metal strip fitted over piezoelectric ceramic/quartz disks. Installation in the roadway requires placement of this assembly at a diagonal to provide complete detection between road shoulders. Based on the diagonal installation, a vehicle with four wheels generates four distinct pulses, whereas a motorcycle only generates two. The system includes the roadway sensors, a multi-channel charge power amplifier, an analog-to-digital converter, and a computer. The computer acquires and analyzes the sensor pulses to obtain classification, speed, and weight of traveling vehicles, and then communicates the information in real time to a data- base housed on a remote server (25).The research team did not expect the system to be in production in time to include it in the field testing for NCHRP Project 08-81. Magnetic Detectors. Magnetometers are passive devices that detect perturbations in the earth’s magnetic field due to the magnetizable components of vehicles as they pass through a detection zone. Magnetometers detect cars and trucks, but they are less effective in detecting and classifying motorcycles and bicycles because smaller vehicles have low magnetizable masses (17, 18).

19 occupancy outputs from selected new detectors against an accurate baseline system, the Peek ADR-6000. Their research tested promising non-intrusive vehicle detector technolo- gies, including video detectors, acoustic, magnetic, inductive loops, and microwave radar (29). Kanhere et al. at Clemson University (2010) tested a track- ing video detection system with their algorithm to count motorcycles at two sites during a motorcycle rally held in Charleston, South Carolina (30). At one site, researchers mounted the video camera in the median of a four-lane divided highway; at the other, they mounted the camera on the side of a road carrying two lanes of traffic in a single direc- tion. Motorcycles accounted for 56 percent of vehicles at the first site and 69 percent at the second. Overall, the Clemson system overcounted motorcycles in the departing direction by 4.4 percent at the first site, under- counted them in the approaching direction by 2.6 percent at that site, and undercounted them by 6.2 percent at the second site. Reportedly, undercounts occurred because of motor- cycles sharing a lane and because of occlusion. The next stage of research involved improving the robustness of the system in those situations, as well as extending the work to handle motorcycles at nighttime and in low ambient lighting condi- tions. Based on the literature, the developers planned further research to augment the algorithm by incorporating pattern- based and shape-based descriptors to better differentiate motorcycles in difficult and ambiguous situations (30). The Clemson system became a market-ready product, sold under the name of TrafficVisionTM, and was eligible for testing as part of NCHRP Project 08-81. Microwave Radar Detectors. Typical mounting struc- tures and locations for microwave radar detectors are poles located adjacent to the roadway or over the lanes to be moni- tored. When vehicles pass through the antenna beam, a portion of the transmitted energy reflects back toward the antenna. The energy then enters the receiver, which detects and calculates the desired data (17). Radar detectors have similar performance characteristics as video-based detectors in terms of occlusion. However, they do not require artificial lighting and microwave radar detectors are virtually unaf- fected by weather (18). As with video, the detection accuracy of radar detectors depends on the truck and motorcycle vol- umes on the monitored roadway or lane(s). At locations with larger volumes of tall vehicles, microwave radar detectors are likely to produce significant undercounts of motorcycles because of vehicle occlusion (18). Advantages of microwave radar detectors include: • Insensitivity to inclement weather at relatively short ranges • Allowance for direct measurement of speed • Coverage of multiple lanes (up to 10 lanes) a tripwire detector. Other video detectors determine when a target vehicle enters the field of view and track the vehicle through this field of view. “Tracking” video systems usually offer more output options but also involve more processing than tripwire systems (26). Video cameras can be mounted on the side of the road, in the median, or directly over the roadway (18). Depending on the type of video detector used, the cameras can collect data about vehicles’ volume, speed, presence, occupancy, den- sity, queue length, dwell time, headway, turning movements, acceleration, lane changes, and classification (26). Video is useful for collecting classification counts of motorcycles and other vehicles but it can only classify based on vehicle length. Detection accuracy depends on the truck and motorcycle volumes on monitored roadways or lanes. At locations with higher volumes of tall vehicles, video detec- tors are likely to undercount motorcycles because of vehicle occlusion. Overhead video cameras do not encounter as much side-to-side occlusion as side-mounted cameras, but they are still subject to front-to-back occlusion. However, their need for large overhead structures for mounting and artificial lighting for nighttime counts may limit their appli- cation in motorcycle detection (18). Advantages of video-based detectors include their ability to monitor multiple traffic lanes and detection zones/lanes, process a rich array of data from multiple cameras, and add or modify detection zones with ease. Major disadvantages of video detectors include: • Vulnerability to viewing obstructions • Inclement weather • Shadows • Occlusion • Light transitions (e.g., day to night) • Vehicle/road contrast • Water, salt grime, icicles, and other debris on the camera lens Also, some video-based detectors are susceptible to distortion from the camera motion caused by strong winds (23). Researchers have investigated various performance aspects of video-based detectors. Middleton and Long- mire (2008) tested a tripwire video system on a road with 10 percent trucks for several hours a day over two 3-day peri- ods, finding that motorcycles were undercounted by about 17 percent (28). The performance aspects of new detectors are improving, making them viable replacements for inductive loops in some cases. Middleton, Parker, and Longmire (2007) conducted research on video detectors to investigate performance aspects of newer detectors. As part of this effort, the research- ers conducted field tests that compared count, speed, and

20 counting rate is about 5 percent, suggesting that TIRTLs may warrant further evaluation for use on two- and three-lane roadways (18). Another detector product using technology similar to the TIRTL is Peek Traffic’s AxleLight laser sensor. The literature search found limited information on this detector and found no sources indicating its accuracy for motorcycles. Research sponsored by the Minnesota DOT (Minge, Kotzenmacher, and Peterson 2010) included the AxleLight but results did not report its accuracy for motorcycle detection (31). Positive attributes of both the TIRTL and the AxleLight include their non-intrusive nature and their ability to detect axles. They are the only known non-intrusive detection systems that can classify vehicles according to FHWA’s Scheme F. However, their negatives include: • High cost (AxleLight retails for $31,580; TIRTL is similar) • Being subject to vandalism and theft (mounted low to the ground) • Difficulty of installation in northern states in winter with snow/ice accumulation alongside the roadway • Potential for increased detection errors in rainy weather due to the spray causing false detections Passive Acoustic Detectors. Passive acoustic detectors can detect volume, speed, occupancy, and classification. They measure the acoustic energy or audible sounds produced by a variety of sources that are generated by a passing vehicle. Sound energy increases when a vehicle enters the detection zone and decreases when it leaves. A detection threshold determines the termination of the vehicle presence signal. Sounds from locations outside the detection zone are attenu- ated (17). Performance tests conducted at TTI indicate that acoustic detectors are not as reliable or accurate as some other non-intrusive detectors (e.g., microwave radar) (29). Literature Summary and Conclusions The literature and Internet search produced several use- ful resources on existing technologies to detect and clas- sify motorcycles. New technologies and sensors continue to emerge as manufacturers respond to the expanding needs of integrated and mobile motorcycle detection systems. No single device is best for all applications. Each detector has strengths and limitations that make them suitable for some purposes but not for others. To that end, the successful appli- cation of detector technologies depends on proper device selection to meet specific needs. Many factors, including data type, data accuracy, installation and calibration, cost, and reliability, impact the selection and performance of detector technology (22). A study of frequency-modulated continuous-wave micro- wave radar systems produced an overall motorcycle under- count of about 19 percent (as opposed to 17 percent for video), with particularly large (and unexplained) undercounts observed on two of the six days of testing (17). Active and Passive IR Detectors. Active IR sensors transmit low-energy laser beams to a target area on the pave- ment and measure the time for the reflected signal to return to the sensor. The corresponding reduction in time for the signal to return indicates the presence of a vehicle. Active IR sensors provide vehicle presence at traffic signals, volume, speed measurement, length assessment, queue measurement, and classification (17). The strength of active IR detectors is that they transmit multiple beams for accurate measurement of vehicle posi- tion, speed, and classification. Also, multiple units can be installed at the same intersection without interference from transmitted or received signals, and multi-zone passive sen- sors measure speed (23). Passive IR sensors detect the energy that is emitted from vehicles, road surfaces, other objects in their field of view, and from the atmosphere, but they transmit no energy of their own (17). Passive IR sensors with a single detection zone measure volume, lane occupancy, and passage. The Transportable Infrared Traffic Logger (TIRTL) is an active IR traffic counting and classification system. Because TIRTLs are designed to detect tires, unlike most other non- intrusive classifiers, they are capable of axle classification instead of length classification. A TIRTL consists of a tire- height transmitter placed on one side of a roadway and a receiver placed on the other side. The transmitter generates two parallel IR beams at a 90-degree angle with the roadway and two additional beams at a diagonal angle. Thus, the four beams consist of two parallel beams and two beams in a criss- cross pattern. The occlusion problem applied to the TIRTL in the con- text of motorcycle detection could be less than that found with length-based classification technologies because of the TIRTL’s detection of tires instead of the entire vehicle. In concept, TIRTLs may be installed permanently for use as continuous counters or temporarily for the collection of short counts. However, it is important that the devices be securely located for protection against vandalism and so that vibrations caused by truck traffic do not cause beam mis- alignment. The accuracy of TIRTL classification counts and TIRTL counts of motorcycles, in particular, varies inversely with roadway width. Testing of TIRTLs on two-lane roadways with an overall width (including shoulders) of 39 ft or less indicates that TIRTLs tend to overcount motorcycles. The overall net over-

21 team began its selection of candidate systems by considering the following existing technologies: • Inductive loop/piezoelectric sensor systems (with full lane- width piezos) • Multi-beam IR sensors such as TIRTL • Microwave radar detectors • Tracking video detectors • Magnetometers However, there were also reasons for looking beyond the tra- ditional technologies. These reasons included the following: • Some detectors involve traffic disruptions during instal- lation. • Some systems do not cover the full lane-width (e.g., some piezoelectric sensors). The existing technologies that have the most promise for accurately classifying motorcycles for long-term classification counts are active IR (TIRTL), inductive loops/piezoelectric sensor systems (with full lane-width piezos), tracking video, and magnetic detectors. Of the prominent magnetic detec- tors, one is intrusive but considered worthy of consideration. Table 5 provides a summary of strengths and weaknesses for these promising technologies based on the literature and Internet review. New and Promising Technologies for Motorcycle Detection Some of the technologies discussed in the previous sec- tion may have been installed for other purposes but are viable for motorcycle detection. Given that resources for purchas- ing new systems may be scarce at the state level, the research Technology Strengths Weaknesses Pneumatic road tubes • Inexpensive • Quick installation for temporary data recording • Easy to maintain • Mature technology • Undercounts motorcycles in groups • Difficulty distinguishing subcompact vehicles and motorcycles • Difficulty detecting very light motorcycles • Motorcyclists intentionally avoid in light traffic • Accuracy subject to weather and traffic conditions • Appropriate only for short-term counts Piezoelectric sensors • Accurately detects motorcycles when new and covers full lane-width • Mature technology • Unpredictable failure Magnetometers • Less susceptible than loops to stresses of traffic • Can be used where loops are not feasible (e.g., bridge decks) • Not sensitive to weather • Single magnetometers undercount motorcycles Microwave radar detectors • Non-intrusive • Quick and easy setup • Side-fire covers up to 12 lanes • Immune to weather and light • Reasonable cost • Undercounts groups of motorcycles • Occlusion causes undercounts Tracking video detectors • Non-intrusive • Reasonably good counts where truck volumes are low • Easy to modify detection zones • Provides view of roadway for verification purposes • Undercounts when truck volumes are high • Requires lighting at night or IR camera • Accuracy reduced during some light and weather events Multi-beam infrared (IR) detectors • Non-intrusive • Scheme F classification by detecting axles • Vandalism and theft potential • Accuracy compromised by weather • Installation causes minor traffic interference • High cost for some sites Acoustic detectors • Non-intrusive • Multiple-lane detection • Insensitive to precipitation • Temperatures affect data accuracy • More time for setup than other systems Table 5. Summary of technology strengths and weaknesses.

22 cation accuracy. Additional research is needed to improve polymer materials for better conductivity and to improve the electronic interface to increase the speed and accuracy of segment closure detection. Additional research on opti- mum segment lengths for other applications not included in the initial field tests (e.g., super single tires, motorcycles, and bicycles) also is needed (34). Independent information from both the developer of the sensor and the Florida DOT in October 2011 indicated that the Florida DOT will continue funding to support further development of the segmented sensor. However, this sensor was not ready for field testing in NCHRP Project 08-81. Modified Radar Detector. In 2010, TTI tested a modi- fied Wavetronix SS-125 High Definition (HD) radar detec- tor, which is designed for motorcycle and bicycle detection (36). Results from a local high speed freeway (SH 6) in Col- lege Station, Texas, provide an indication of the HD detec- tor’s performance during heavy rain on 1 day compared to a similar period on the following day with no rain. Its overall count accuracy for all vehicles during this 50-minute period was 99.4 percent both during the heavy rain and 99.4 per- cent during a dry period of the same length. An evaluation to determine the number of motorcycles that would be detected assuming length bins of less than 10 ft and less than 8 ft (with the latter bin intended to contain only motorcycles) revealed that lanes nearer the HD had greater propensity to overcount vehicles of the selected lengths. Results from a 2010 study at one of TTI’s controlled test facilities indicate that detection accuracy of motorcycles in a staggered pattern was 96.1 percent, and the average speed difference between the HD and a baseline GPS unit was 0.33 mph with a standard deviation of 1.49 mph. Similar results were obtained with motorcycles side-by-side as long as they were separated by about one lane-width (36). The research team made repeated attempts to contact the manufacturer of this radar detector without success. The FHWA contact on the SBIR test pertaining to this sensor recently divulged that the manufacturer had declined the opportunity to enter Phase 2 testing. Given that the current modified detector had been tested already, the research team did not pursue further tests in NCHRP Project 08-81. Wireless Magnetometers. Early tests using single wire- less Sensys Networks magnetometers for detection of a mix of vehicles resulted in reasonable performance, which has since improved. Since these early tests, the manufacturer has added sensitivity settings that should improve the detector’s motorcycle detection. The manufacturer also has recom- mended installing two or three detectors in a side-by-side configuration to increase the detection area of these detec- tors, making it less likely that motorcyclists can avoid the detection area. • Some technologies have not been sufficiently tested for motorcycle detection. Promising technologies for motorcycle detection should be able to address several detection and classification issues that are unique to motorcycles. These issues include the abil- ity to distinguish motorcycles from subcompact cars and the ability to count motorcycles in groups. The combination of one or more detector systems (multi-technology) may better address motorcycle detection where size is critical. Follow- ing are some salient points regarding new technologies that might be candidates for field testing. One of the options is a viable multi-technology system. Some of the technologies discussed in the next sections were not included in the testing conducted for NCHRP Project 08-81 but may merit addi- tional research in the future. Inductive Signature Technologies. As part of a project sponsored by the Arizona DOT, TTI had requested informa- tion on an alternate ground truth system offered by Inductive Signature Technologies, Inc. (IST) (32).At that time, some uncertainty existed about the IST system’s perceived lack of maturity compared to another system using similar technol- ogy. Since that time, the IST system has been involved in an FHWA-sponsored research project on length-based classifi- cation (33). However, the IST system is still not market ready as it requires special knowledge for interpretation of results and special skills to install the system in the roadway. For these reasons, the research team elected not to test the IST system as part of NCHRP Project 08-81. Segmented Piezoelectric Axle Sensor. Recent tests of a segmented sensor for high resolution detection show high potential for detection and classification of motorcycles. However, this sensor is still in the development stage, with the Florida DOT sponsoring early proof-of-concept tests of prototypes (34). The concept involves multiple sensing ele- ments housed in a single long channel. The finished product will probably look like a standard piezoelectric sensor in most ways, with the exception that along its length it will have sev- eral independent sensing elements. The length of each sensing element will determine the accuracy of the sensor in establish- ing tire widths, tire separation in dual tire configurations, and detecting motorcycles. Early prototypes were 8.0 ft in length. Researchers used a modeling approach to determine opti- mum segment dimensions. They varied segment lengths from 0.5 in. to 4.0 in. in increments of 0.05 in. Results indicated that a length of less than 0.9 in. can result in 100 percent discrimi- nation between dual and single tires. The research report did not comment on whether this length would be feasible (35). Field test results of the segmented sensor indicated that it can differentiate between single and dual tires and that such distinction can significantly improve vehicle classifi-

23 2. They used the IR thermal camera to distinguish cars and motorcycles from bicycles. 3. They used the acoustic sensor to distinguish classes of motorcycles (37). The first step in the detection process is to use the IR ther- mal camera to distinguish bicycles from other vehicles, keying on the unique thermal signature of bicycle wheels. Next, the distinction between motorcycles and cars comes from their different thermal signatures, with each exhibiting its own shapes and characteristics. The operator can select a subarea of the image and look at its thermal threshold and shape to determine the vehicle that generated the image. Because the IR camera technique sometimes confuses cars with motor- cycles, the system offers a second source of input—the visible light (stereo) camera. Its primary source of input is the vehi- cle wheels. To distinguish motorcycles from cars, it searches the shape of the area around the front wheel. For a motor- cycle, the front wheel area is distinct and different from that of a larger vehicle (37). At this point in the classification routine, only vehicles clas- sified as motorcycles remain. The next step is to determine if the vehicle is a motorcycle, a moped, or a scooter. These three vehicle types have distinct sound signatures, so the system uses the acoustic detector to distinguish between the three vehicle types. The research team uses digital signal processing with phase analysis of the sound to measure the spectrum features of the vehicle classes. Migma researchers conducted an outdoor test as the Phase 1 project development concluded to determine the accuracy of the multi-technology system. Table 6 shows the results, indicating some classification errors. Out of 12 cars, the system classified one as a heavy motorcycle and one as a light motorcycle. Out of 14 heavy motorcycles, the Migma system classified three as cars. Out of four light motorcycles, it classified one as a car; and of The most recent evaluation of Sensys Networks magne- tometers (prior to NCHRP Project 08-81) occurred at the TTI test facility in College Station in 2011, although this test still did not specifically target motorcycle detection. The test- ing used three physical magnetometer configurations and various extension times to determine the best count accu- racy. This research used contact closure cards and a vehicle classifier to receive input from the magnetometers and com- pared detections to a high end classifier using inductive loop signatures. Findings indicate that a single wireless magne- tometer sensor can produce excellent count accuracy for the full range of vehicle types when the proper extension time is used (especially with a significant number of large com- bination trucks). The research explored the recommended settings, finding that when the proper extension time was implemented, wireless magnetometers provided consistent count accuracy greater than 99 percent when compared to the accurate baseline system. Based on early tests and a local representative who could provide technical support, this detector was selected as a viable candidate technology for NCHRP Project 08-81. Multi-Technology System. Prior to its testing in NCHRP Project 08-81, an earlier test in 2011 of a new multi-technology system produced by Migma Systems, Inc. also showed prom- ise to accurately detect motorcycles, although it was not in full production during the research timeframe of NCHRP Project 08-81. With funding through an SBIR project, Migma Systems, Inc. developed a multi-technology system using an IR visible light stereo camera, an IR thermal camera, and an acoustic sensor. Each of these components had a specific detection role to play. The developers designed the system around these three components: 1. They used the IR camera to identify the riders on two- and three-wheeled vehicles. Vehicle Class Bicycle Heavy Motorcycle Light Motorcycle Car Moped Scooter Bicycle 8 0 0 1 0 0 Heavy Motorcycle 0 11 0 3 0 0 Light Motorcycle 0 0 3 1 0 0 Car 0 1 1 10 0 0 Moped 0 0 0 0 3 0 Scooter 0 0 1 0 0 2 Source: Reference (37). Table 6. Summary of classification results from Migma detection system.

24 • The registered owner may not be the person riding a motorcycle. • At least one state offers lifetime motorcycle registration. The U.S.DOT re-baselined its motorcycle fatality rate mea- sure for FY 2008 to reflect a change of focus from fatalities per 100 million VMT to fatalities per 100,000 registrations (39). VMT is considered the best measure for exposure because it measures actual miles traveled and is U.S.DOT’s preferred method of measuring fatality rates. Measures of effectiveness of highway safety trends must include exposure to be meaning- ful. For example, fatality rates based on population, the num- ber of registered vehicles, or the number of licensed drivers are of little value as analytical tools due to lack of exposure data. Motorcycle Sales Data. Data are available on the num- ber of motorcycles sold in the United States every year. These data help provide a rough estimate of the number of motor- cycles that may be on the road, but they are not particu- larly helpful in understanding the number of miles ridden or demographic data on riders. For example, motorcycle sales data do not reflect miles ridden on older motorcycles or miles ridden by non-owners. Also, many motorcyclists own more than one motorcycle, which potentially skews the data. Motorcycle License Data. All states require motor- cyclists to have a separate motorcycle license or endorse- ment. However, research shows that a high proportion of riders have not obtained motorcycle licenses. There is also possibly an equally high population of licensed motorcycle riders who do not ride. Highway Usage Data. States are required to keep data on the number and nature of vehicles using major highways. These data may give some insight into the number of motor- cycles on these roads and the numbers of miles being rid- den, but the data might not come from highways where most motorcyclists are traveling. Travel Demand Data. States maintain data on the extent of road usage so that they can make forecasts of future needs and they can determine the effects of traffic on the environ- ment. These data may not contain information about how much of this traffic consists of motorcycles. Surveys of Motorcyclists. It may be possible to deter- mine information on motorcycle travel by surveying motor- cyclists. The advantage of this approach would be gathering a large amount of information on riders. The disadvantages would be the same disadvantages of survey research discussed above, plus the fact that all data would be self-reported and nine bicycles, it classified one as a car. It classified all three mopeds correctly, but it classified one of three scooters as a light motorcycle (37). The Migma research team identified reasons for the inac- curacies and has worked to improve the weaknesses. The team is continuing this refinement in the SBIR Phase 2 proj- ect, which is currently underway. Phase 2 includes testing the system with motorcycles traveling in a group and testing in inclement weather. Given that Migma Systems, Inc. promised to provide a detection system and technical support during the field tests for NCHRP Project 08-81, the research team included this multi-technology system in field testing. Estimating Motorcycle VMT Six states currently use seasonal and day-of-week factors derived solely from continuous counts of Class 1 vehicles to convert short counts of Class 1 vehicles to estimates of AADT. Because the use of Class 1 vehicles is subject to substantially greater seasonal and day-of-week variation than the use of vehicles in other classes, use of factors derived solely from continuous counts of Class 1 vehicles is usually a prerequisite for producing reasonable AADT estimates for these vehicles. One additional state was expected to start deriving factors solely from continuous counts of Class 1 vehicles in 2009, and two more plan to do so in the future (38). Ten states develop VMT estimates for Class 1 vehicles without using factors derived solely from counts of Class 1 vehicles. These states did not indicate any plans for changing this procedure. Another option for producing improved esti- mates of motorcycle crash and fatality rates is to implement procedures for estimating the VMT of Class 1 vehicles that do not require vehicle counts (18). Additional sources are available from which to obtain data on the number of motorcyclists riding on which roads. This section summarizes current methodologies for calculating incident and fatality rates when the available motorcycle volume and VMT data are less than comprehensive. It also covers adjustment factors that are available for use in related calculations. Motorcycle Registration Data. Registration data con- tain demographic data and data about motorcycles. However, there are issues that obscure the connection between registra- tion data and exposure data. These issues include: • Some motorcycles may not be registered (though unreg- istered motorcycles are probably untagged and therefore not ridden on public roads). • Some riders own and register multiple motorcycles (but can ride only one at a time).

25 cessfully) to detect motorcycles. DOTs in the following states were surveyed: • Arizona • California • Colorado • Florida • Minnesota • New York • Ohio • Oregon • Virginia • Washington Technology Used by State DOTs Arizona The Arizona DOT found that using inductive loops and piezoelectric sensors in combination is effective in detecting motorcycles in most cases. The rating that the Arizona DOT personnel gave for other technologies (inductive loops alone, magnetometers, and radar) was “poor.” For video, their rating was “good with manual classification,” apparently meaning that they recorded video and conducted a subsequent off-line manual classification. The Arizona DOT has experienced an average life expectancy of the loop/piezo system of 7 years. Loops cost the Arizona DOT about $1000 installed and piezos cost about $900 each (length not specified). Assuming two piezos and one loop per lane, the system installation cost would be about $2800 per lane. The Arizona DOT included some comments about the strengths and weaknesses of each technology. The strength of the inductive loop was that it works well in all traffic con- ditions. Its weakness is that it needs piezoelectric sensors to detect motorcycles. A strength of an older magnetometer is that the user can adjust the sensitivity, but the weakness is the high cost of installation (which includes boring under the roadway). The strength of piezoelectric sensors is their accuracy, but their weakness is that they need inductive loops for presence detection. California The California DOT (Caltrans) has considered a number of different technologies to detect vehicles in general and motorcycles in particular. Technologies used by Caltrans have included: • Inductive loops • Piezoelectric sensors • Magnetometers self-reported exposure data may not be accurate. Potential surveys include: • National Household Traffic Survey (NHTS): This survey samples households by telephone (including cell phones) from all regions across the country. The data collected include demographic characteristics of households, people, vehicles, and detailed information on travel for all purposes by all modes (5 percent owned motorcycles, 3.6 percent of all vehicles were motorcycles) (40). • Origin and destination (O & D) surveys: These surveys gather travel information and often are used by agencies to determine future traffic patterns. • Driver exposure surveys: These surveys examine the rela- tive safety of the road transport system by asking drivers to provide information about distance and duration trav- eled as well as “opportunities for accidents”—the drivers’ exposure to the possibility of accidents. Roadside Counts. One way to determine the number, and to some extent the nature, of motorcycles on a particu- lar roadway is to manually count motorcycles through visual observation. The results can include estimates of motorcycle size and type (e.g., sport bikes, cruisers) and determination of riders’ helmet status (e.g., type of helmet or no helmet). Helmet type (e.g., no helmet, novelty helmet, or full-face hel- met) could determine whether the observer could determine driver age, race, and gender. Recorded Video from Traffic Cameras. Many states and local jurisdictions have traffic cameras that are capable of recording traffic for subsequent analysis. This option would be similar to roadside counts except that the camera locations might not be optimum for capturing motorcycle travel, and the orientations and positions of the cameras might preclude capturing certain critical details. Insurance Company Data. Insurance companies keep records that may contain information on number of motor- cycles in a household, number of riders, and percentage of use of each motorcycle by each rider. Data would also include demographic data and may include crash data. However, insurance companies do not generally make data available for research purposes, and the data are often not sufficiently detailed for research purposes. Agency Engagement The information in this section comes from a survey of 10 state DOTs that responded to questions about the technol- ogies and methodologies they used (successfully or unsuc-

26 cycles, although there had been no formal evaluation specifically targeting motorcycles. The Florida DOT loops are spaced 16 ft apart lead to lead with a 6-ft piezoelectric sensor between the two loops covering one wheel-path. Even though the Florida DOT had not tested this system for motorcycle detection, their spokesman said it “seems to detect most of them if they do not intentionally avoid the sensors.” According to the Florida DOT spokesman, the problem with this system in detecting motorcycles is in classification since the number of small cars with a similar wheelbase has increased. The inductive loop/piezoelectric sensor system cannot distinguish between them, so there is a fair amount of misclassification. The Florida DOT also uses pneumatic tubes for short-term counts, and they appear to detect almost everything that crosses them. The Florida DOT has investigated other technologies such as video, radar, and magnetometers. The radar was not very accurate for vehicle classification although the Florida DOT did not conduct a formal test of its accuracy for motor- cycles. It overestimated the number of trucks by a large fac- tor at one site over a 1 month period. The Florida DOT also tested wireless magnetometers placed two per lane, and they counted accurately but could not classify length well enough to meet the state DOT’s needs. The same was true of Mio- Vision video in a Miami test—the video counted well, but it was not sufficiently accurate at classification. Life-cycle costs are a function of the life of each component of the system (loops and piezos). The Florida DOT experi- ences about a 5 to 6 year life with piezoelectric sensors. Induc- tive loops usually last twice as long—10 to 12 years. The major factor in the life of in-pavement systems is the care during installation used by the installation contractor. The Florida DOT did not provide information about actual costs, just the time they last. The actual installation time for the inductive loop/piezoelectric sensor system is about 2 hours, but adding in traffic control setup and removal and epoxy cure time plus loop sealant cure time takes a full day for a two-lane site. The Florida DOT has not found a satisfactory non-intrusive detection system to cover all vehicle types. For that reason, the agency continues to use inductive loops/piezoelectric sensors for long-term counts and pneumatic tubes for short- term counts. Minnesota The Minnesota DOT uses one non-intrusive detector for motorcycle detection: side-fire radar. The Minnesota DOT also uses two intrusive systems for motorcycle detection—one with piezoelectric sensors and loops and the other with quartz weigh-in-motion sensors and inductive loops. The Minnesota DOT also uses pneumatic tubes for short-term counts. • Radar • Video Caltrans has recently created a policy of minimizing the presence of personnel on the roadway for installation and maintenance of detectors. Caltrans engineers are actively evaluating newer technologies with an emphasis on non- intrusive technologies while maintaining accuracy, cost, and ease of use. Based on experience, Caltrans has found that a combination of inductive loops and piezoelectric sensors provides the most accurate and most cost-effective technol- ogy to count vehicles and motorcycles. Some districts within Caltrans are experimenting with magnetometers. However, details pertaining to their accuracy were not available. Cal- trans has encountered numerous issues with radar accuracy for counting motorcycles, including occlusion of motorcycles due to larger vehicles. Caltrans also has found video detection to be expensive and not very accurate. Like other states, the objective of the Caltrans vehicle monitoring program is to collect data on all vehicles, including motorcycles. In addition, Caltrans routinely sets up temporary detector stations using pneumatic tubes at locations with observed higher proportions of motorcycles. These temporary count stations typically include 1 week of data. The primary metrics used to select technologies are accu- racy, cost, ease of setup, and vendor support. However, the agency has not conducted a formal analysis of the accuracy of technologies for detecting motorcycles. The cost of instal- lation and maintenance of a detector station is an important factor for Caltrans. Caltrans spends approximately $50,000 for an inductive loop/piezoelectric sensor system on four lanes, which typically requires 1 night to install. If the instal- lation is done properly, the only required maintenance is an occasional calibration of the counter. Colorado The Colorado DOT has found that inductive loops alone are not accurate in detecting motorcycles, so the Colorado DOT uses piezoelectric sensors along with loops to improve motorcycle detection. Evaluation metrics were accuracy and cost, supplemented by the fact that the Colorado DOT already uses piezos and loops for volume and classification data collection (for all vehicles). The Colorado DOT has not formally tested the accuracy of this system or other detection technologies for motorcycles. Florida The Florida DOT considered inductive loops and piezo- electric sensors to be the most accurate for detecting motor-

27 Oregon The Oregon DOT uses inductive loops and piezoelectric sensors for motorcycle detection and has found that it pro- vides acceptable results, verified by video recordings. Whereas some states have reported using piezos covering only one wheel-path, the Oregon DOT uses a longer piezoelectric sen- sor to cover the full lane-width. Oregon DOT personnel did not report either the life-cycle cost or the initial cost of any technologies. Of other technologies that the Oregon DOT investigated, radar required less time in the road, which repre- sented a safety and mobility improvement. Also, radar handles lane changes better than other technologies. Virginia The Virginia DOT has found that inductive loop and piezo- electric sensor systems are accurate at detecting motorcycles. However, the Virginia DOT has developed an improved sys- tem for collecting classification data as well as a post-processing methodology to improve detection of light vehicles. Starting on the data collection side, the Virginia DOT developed a specification for a high performance inductive loop board. The loop board is a 4-channel board, which reduces crosstalk by scanning channels. The manufacturer described the better performance of this board as follows: “the detector puts out a higher signal on the loops to increase signal-to-noise ratio.” Also, the Virginia DOT has pretty exacting standards for installing loops that likely contribute to better performance. Components of these standards include: • All loops installed with 4 turns of wire • No wire splices allowed • Wire meets International Municipal Signal Association (IMSA) 51-7 specification • 4 in. deep installation (primarily to survive milling) Similar to the inductive loop board, the Virginia DOT’s piezoelectric sensor board is not an off-the-shelf item. Several years ago, the Virginia DOT worked with a supplier to gather a large number of waveform signals that might typically be found in Virginia. They developed a piezoelectric card with the capability to analyze complex waveforms rather than one that was simply a threshold detector. This board can handle a wide voltage range and also rejects adjacent lane energy. The Virginia DOT’s piezoelectric sensor installation stan- dard is different from that of some states. The Virginia DOT does not use single wheel-path installations—it only uses full lane-width piezos. The Virginia DOT tried several different approaches (e.g., high output piezos, various grout materials) to improve the longevity of the installation. Current practice involves the installation of two piezoelectric sensors per lane, stacked in a single sawcut. The stacked configuration places a The Minnesota DOT has found that the radar detectors are very good for motorcycle detection in free-flow conditions. Road tubes, piezoelectric sensors, and sensors using quartz tech- nology are only accurate if they cover the full lane-width. For cost of each detector type, the Minnesota DOT specified a first cost and a cost per year to maintain. The initial cost of the radar is $6000, with an additional cost of about $600 per year. Road tubes cost the Minnesota DOT $2000 initially plus $200 per year, piezos cost $1200 per lane and $200 per year, and quartz sensors cost $25,000 per lane and $5,000 per year per lane. A negative feature of radar is the need for a bucket truck for installation and maintenance, plus it does not detect as well in stop-and-go traffic. Occlusion is also a factor that affects its accuracy for motorcycles and some other vehicles. A negative feature of road tubes is they are not appropriate for high vol- ume sites due to risk to installers. A problem with piezo electric sensors and quartz sensors is their higher cost and they miss motorcycles if the motorcycles are near the centerline. New York The New York State DOT uses inductive loops with piezo- electric sensors for long-term counts involving motorcycles and two pneumatic tubes with portable counter for short-term counts. The agency’s desired accuracy is 95 percent, and the state believes their detectors achieve that accuracy. The New York State DOT spends about $40,000 for each two-lane clas- sification site using inductive loops and piezoelectric sensors. Ohio The Ohio DOT reported that the LPL configuration has been the mainstay for its overall permanent classification pro- gram. The Ohio DOT found inductive loops and piezoelectric sensors arranged in this configuration to be the most reliable and accurate sensors for collecting vehicle classification data. According to an Ohio State University (OSU) research project conducted at one of the Ohio DOT’s permanent count sta- tions, the LPL counter properly recorded motorcycles 96 per- cent of the time. OSU conducted the study over a 4-hour period on one lane using a video recorder and manual data verification. This was part of an overall classification study at one permanent count station in Columbus, Ohio. The Ohio DOT reported that a LPL site currently costs approximately $10,300 per lane to install. To reinstall sensors only, the cost is approximately $4,900 per lane. Depending upon pavement conditions, sensors need to be replaced every 4 to 5 years. The strength of this technology is that it is fairly accurate across all vehicle classes. A weakness of this technol- ogy is that it fails quickly mainly due to pavement conditions. In addition, it is expensive to install, requiring lane closures to install and/or maintain.

28 degrades. With the old failing sensor (prior to the swap in April 2008), significantly more vehicles were recorded as Class 21 vehicles than afterward with the new sensor. Basi- cally, Class 1 vehicles replaced Class 21 vehicles. In either case, the Virginia DOT system generally counted the motorcycles correctly as either Class 1 or Class 21. The cost of the Virginia DOT system is about $20,000 for a two-lane installation of two inductive loops and one piezoelectric sensor per lane. The Virginia DOT reports that properly installed loops can last indefinitely, while piezo sensors have a shorter life. The Virginia DOT reports a cost of $9,000 (in the context of a two-lane system) for replac- ing piezoelectric sensors, which is considerably higher than other states reported. Finally, the Virginia DOT emphasizes that, even with the implemented enhancements, motorcycles remain a difficult class to monitor. Ridership patterns (e.g., weekends, fair weather) increase the difficulty of this class even when equip- ment works optimally. piezo at 2.5 in. below the surface and another at 1.5 in. below the surface in the same sawcut. The Virginia DOT results to date have been very good, with optimum motorcycle classifi- cation accuracy as high as 98 percent. With regard to the data processing component, a basic understanding of the Virginia DOT system is essential. In the early morning hours of each day, the Virginia DOT autopolls its count stations and uploads the data to its database. During the upload of data, its system completes a number of automated checks based on established performance criteria. It outputs informational and error messages based on these checks. Upon arriving at work later that morning, the Virginia DOT’s con- tinuous count station (CCS) data analysts review the messages from previous day(s) and assign a quality rating to the data. Some messages relate specifically to the settings, programming, and maintenance of the loop. Such messages may also trigger specific actions, such as an information service call to a contrac- tor, a site visit, or scheduling and prioritizing repair of a count station. With the assistance of vendors, the Virginia DOT devel- oped a unique loop/piezo system for monitoring motorcycle travel. It involves a 21 bin table for capturing motorcycles as degraded piezoelectric sensors start missing hits of lighter vehicles like motorcycles (see Table 7). Even though it works well for most detections of motorcycles, the system still misses Class 1 vehicles that straddle lanes or that ride side-by-side or in tight packs. The Virginia DOT also stated that there are issues to be addressed in concrete pavement with significant rebar and missed detections of smaller motorcycles (mopeds, scooters, etc.). In general, the Virginia DOT has observed that its system accurately classifies full-size motorcycles as Class 1 or Class 21, depending on the piezo signal. Figure 10 indicates how the Virginia DOT results change with replacement of a failing piezoelectric sensor, indicat- ing that motorcycles are detected even as the sensor’s signal Vehicle Class Description 1 through 15 Standard vehicle classification 16 and 17* 1 axle detected Class based on magnetic length 18 and 19* Zero axles detected Class based on magnetic length 20 Zero or 1 axle detected Magnetic length over 22 ft 21 Zero or 1 axle detected Magnetic length less than 7 ft * Equivalent to FHWA (Scheme F) Class 2 and Class 3. Table 7. The Virginia DOT 21 bin classification table. Piezo Replacement Figure 10. The Virginia DOT classification results change with new piezoelectric sensor.

29 77 Accurate counts in areas where there is a lot of lane shifting – Cons: 77 Poor vehicle lengths 77 Speeds just averaged 77 No axle classification • Video – Con: 77 Does not meet minimum standards for vehicle count and classification data Table 8 summarizes the technologies used by state agencies that were contacted by the research team. In some cases, state DOTs reported on technologies not currently used but eval- uated using non-formal procedures. The comments gener- ally emphasize motorcycle detection but also consider other vehicle types. Methodology Used by State DOTs The Arizona DOT submits motorcycle data to meet the HPMS requirement by aggregating motorcycle data on a statewide level. The only other comment that Arizona DOT provided about this reporting was that they “. . . count [motorcycles] based on general observations as we do with other vehicles.” Alternatively, state DOTs have the option of reporting disaggregated data by highway class, urban area (or urban/ rural), roadway segment, or other. Agencies that choose dis- aggregate reporting have the opportunity to describe how they determine locations to monitor motorcycle travel. The Caltrans data collection group collects the data at designated count stations, runs it through data checks, and provides the data to the HPMS group in Caltrans. These data come from both the permanent count stations and the tem- porary count stations. HPMS then applies certain factors to the data to generate motorcycle counts for reporting purposes. The Colorado DOT monitors motorcycle travel at the same locations where permanent count stations are located and where 24- to 48-hour counts are available. The Florida DOT investigated and established the sites for all traffic data collection about 20 years ago (not specifically target- ing motorcycles). The state has 300 CCSs to cover 12,000 miles of roadway across the state. The Florida DOT wants at least one classification site on each roadway section per county, although some districts have more than the minimum. The Minnesota DOT counts motorcycles at the same classification sites as other vehicles. The detection equip- ment used is piezoelectric sensors and loops or quartz sensors and loops. Washington State The Washington State DOT has investigated using induc- tive loops, magnetometers, radar, and piezoelectric sensors. The metrics used to evaluate the sensors include: • Accuracy (using manual verification counts). • Cost (compared to similar equipment). • User interface: The DOT ensures that the equipment inter- faces well with its own software and checks hardware com- patibility as well. The agency wants its technicians to be able to learn the system quickly. • Vendor support: The DOT needs to easily contact the ven- dor, to get questions answered quickly, to get quick turn- around on hardware orders, and to have materials delivered quickly. The Washington State DOT found that radar was reason- ably good for motorcycle detection but that its classification accuracy was not as good. Magnetometers and video were less accurate, but inductive loops and piezoelectric sensors could be excellent with proper setup. Life-cycle costs were not available. The initial cost of an inductive loop was $1,600, and a loop typically lasts from 5 to over 30 years. The initial cost of a piezoelectric sensor was $1,200, and a sensor might last 3 to 5 years. According to the Washington State DOT, the pros and cons of each technology are as follows: • Inductive loops/piezoelectric sensors – Pros: 77 Proven technology 77 Very accurate speed and length classification 77 Inexpensive installation costs 77 Loops can be used in conjunction with piezoelectric sensors for 13 bin axle classification – Cons: 77 Static position 77 Dependent on road conditions • Magnetometers – Pros: 77 Quick and easy installation 77 Reasonably accurate counts 77 Inexpensive to add lanes – Cons: 77 Poor interface with customer-supplied equipment 77 Poor speed data 77 Poor length data 77 No 13 bin classification • Radar – Pros: 77 Non-intrusive technology 77 Accurate when using higher level equipment

30 Agency Technology Pros Cons Arizona DOT Inductive loops/piezoelectric sensors Video with manual classification All others Good in all traffic conditions Good (with manual classification) N/A N/A Poor by itself Poor Caltrans* Inductive loops/piezoelectric sensors Radar Most accurate detector Cost effective N/A Poor motorcycle detection Colorado DOT* Inductive loops alone Inductive loops/piezoelectric sensors N/A Accuracy and cost acceptable Poor motorcycle detection Florida DOT* Inductive loops/piezoelectric sensors (sequence: LPL) Radar Other non-intrusive sensors Most accurate detector Count accuracy acceptable N/A Cannot distinguish subcompacts Class accuracy poor None found acceptable Minnesota DOT Radar Road tubes Piezoelectric sensors Quartz WIM** Accurate in free-flow OK if full lane-width OK if full lane-width OK if full lane-width Need bucket truck to install Short-term counts only N/A N/A New York State DOT Inductive loops/piezoelectric sensors Road tubes (two) Use for long term Use for short term N/A N/A Ohio DOT Inductive loops/piezoelectric sensors (sequence: LPL) Primary for all class counts 96% accurate for motorcycles N/A Oregon DOT Inductive loops/piezoelectric sensors (full lane-width) Acceptable accuracy for motorcycles N/A Virginia DOT Inductive loops/piezoelectric sensors (full lane-width) -High performance loop board -High output piezos -21 bin class table Highly accurate for motorcycles N/A Washington State DOT Inductive loops/piezoelectric sensors Radar Magnetometers Video Accurate, low cost Accurate if high level equipment Accurate for lane changing Accurate counts, easy to install N/A Static position, longevity a function of pavement condition Poor vehicle lengths and speeds Poor vehicle lengths and speeds Poor length and class * No formal analysis conducted specifically for motorcycles. **WIM = weigh-in-motion. Table 8. Summary of technologies and assessments by state agencies.

31 For the transportation agencies at these four states, calcu- lation of the data collection protocols used correlation coeffi- cients (Pearson’s R) to measure the association between crash frequency (unweighted and weighted) and traffic volume (motorcycle and total traffic) at the count stations nearest the crashes. Calculation of weighted crash frequency is done using Equation (1), which appears in Chapter 2 of this report as part of the discussion of the methodology used to deter- mine the correlation coefficients. Michigan Analytical Results Figure 11 and Table 9 show the correlation coefficients (Pearson’s R) for Michigan data measuring the association between crash frequency (unweighted and weighted) and traffic volume (motorcycle and total traffic) at the count sta- tions nearest the crashes. Weekday and Weekend Motorcycle Crashes in Michigan. Figure 12 and Table 10 show the Michigan crash frequency and volume count correlations separately for weekday and weekend crashes. The New York State DOT collects data for HPMS on a statewide aggregated level. The agency estimates statewide motorcycle travel data based on growth factors. The Ohio DOT reports motorcycle travel on a disaggre- gated level by highway class and by urban or rural categories. The Ohio DOT does not currently select sites based specifi- cally on motorcycles, but uses the statewide data collection plan based on roadway groupings. The Oregon DOT collects enough samples and breaks state roadways into enough segments to satisfy its own decision-makers. The agency also collects data on non-state roadways, but it does not specifically target motorcycles. The Oregon DOT counts motorcycles based on general observa- tions as it does with other vehicles. The Oregon DOT does not go to extra lengths to obtain motorcycle data because the benefits would not be worth the additional cost. The Virginia DOT reports HPMS data in disaggregate form by roadway segment. The agency breaks every road into traffic links. On all roadways functionally classified as collector and above, the Virginia DOT includes sampling sections of collected classification data. Site selection is based on locations that are perceived to be best for all vehi- cles. The Virginia DOT collects motorcycle data at almost all locations where classification data are needed using the LPL configuration for continuous count sites and pneu- matic tubes at the remaining sites (about one-third of the total sites). The Washington State DOT does not report motorcycle data from its short count program. In its permanent count program, the agency counts motorcycles based on general observations as it does with other vehicles. The Washington State DOT collects data 365 days a year from 162 perma- nent locations and reports the data to FHWA by hour and by direction. None of the state agencies contacted uses a special pro- cedure for locating the motorcycle classification count loca- tions. They consider motorcycles like all other vehicles and count them on the same basis. Some state DOTs aggregate motorcycle data on a statewide basis and report data by high- way class and/or urban versus rural, and so forth. No agencies contacted offered special spatial or temporal considerations for motorcycles. This finding led the research team to use crash data to provide the necessary guidance on data collec- tion protocols. Crash Data Collection Protocols The research team obtained information on crash data collection protocols from Michigan, Montana, Texas, and Wisconsin. Traffic Volume Counts Motorcycle Crash Frequency Sample Size (N)Unweighted Weighted Motorcycle 0.266* 0.436** 101 All 0.332** 0.521** 101 *p < 0.005; **p < 0.001 0 0.1 0.2 0.3 0.4 0.5 0.6 Unweighted Weighted Pe ar so n' s R Correlation Category All Motorcycle Figure 11. Correlation of motorcycle crash frequency and the traffic volume at the nearest count station to the crash (Michigan data). Table 9. Correlation of motorcycle crash frequency and the traffic volume at the nearest count station to the crash (Michigan data).

32 Time Period Crash Frequency Traffic Volume Counts Sample Size (N) Motorcycle All Weekday Unweighted 0.302* 0.387* 51 Weighted 0.467** 0.559** Weekend Unweighted 0.279* 0.333* 50Weighted 0.462** 0.552** * p < 0.05; ** p < 0.001 Table 10. Weekday and weekend crash frequency (weighted and unweighted) correlated with motorcycle and total vehicle counts at the nearest count station (Michigan data). 0.0 0.1 0.2 0.3 0.4 0.5 0.6 Unweighted Weighted Pe ar so n' sR Correlation Category All Motorcycle Figure 13. Correlation of motorcycle crash frequency and the traffic volume at the nearest count station to the crash (Montana data). Vehicle Type Motorcycle Crash Frequency Sample Size (N)Unweighted Weighted Motorcycle 0.045* 0.324** 118 All vehicles 0.020* 0.229** 289 *p > 0.05; **p < 0.001 Table 11. Correlation of motorcycle crash frequency and traffic volume at the nearest count station to the crash (Montana data). Montana Analytical Results Figure 13 and Table 11 show the correlation coefficients (Pearson’s R) for Montana data measuring the association between crash frequency (unweighted and weighted) and traffic volume (motorcycle and total traffic) at the count sta- tions nearest the crashes. Texas Analytical Results Figure 14 and Table 12 show the correlation coefficients (Pearson’s R) for Texas data measuring the association between crash frequency (unweighted and weighted) and traffic volume (motorcycle and total traffic) at the count stations nearest the crashes. Wisconsin Analytical Results Figure 15 and Table 13 show the correlation coefficients (Pearson’s R) for Wisconsin data measuring the association between crash frequency (unweighted and weighted) and 0 0.1 0.2 0.3 0.4 0.5 0.6 unweighted weighted unweighted weighted Weekday Weekend Pe ar so n' sR Correlation Category All Motorcycle Unwei W i ted Un i W i ted Figure 12. Weekday and weekend crash frequency (weighted and unweighted) correlated with motorcycle and total vehicle counts at the nearest count station (Michigan data). traffic volume (motorcycle and total traffic) at the count sta- tions nearest the crashes. Weekday and Weekend Motorcycle Crashes in Wis- consin. Figure 16 and Table 14 show the Wisconsin crash frequency and volume count correlations separately for weekday and weekend crashes. Interpretation The purpose of this analysis was to determine if motorcy- cle crash locations are distributed geographically in a pattern that reflects the geographic distribution of traffic volume. The researchers first approached this analysis based on an ini- tial mapping of crash and traffic volume data for the state of Michigan, which showed on casual observation that the two seemed to track well geographically. The goal of this analy- sis was to determine to what extent a state DOT might be able to rely on the spatial distribution of motorcycle crashes when attempting to determine where best to count motor- cycle traffic. It was also important to identify an analysis that most states could perform in making these siting decisions for improved motorcycle count installations.

33 0 0.1 0.2 0.3 0.4 0.5 0.6 Unweighted Weighted Pe ar so n' sR Correlation Category All Motorcycle Figure 14. Correlation of motorcycle crash frequency and the traffic volume at the nearest count station to the crash (Texas data). Vehicle Type Motorcycle Crash Frequency Sample Size (N)Unweighted Weighted Motorcycle 0.253* 0.485* 545 All vehicles 0.193* 0.505* 545 *p < 0.001 Table 12. Correlation of motorcycle crash frequency and traffic volume at the nearest count station to the crash (Texas data). 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 Unweighted Weighted Pe ar so n' s R Correlation Category All Motorcycle Figure 15. Correlation of motorcycle crash frequency and the traffic volume at the nearest count station to the crash (Wisconsin data). (a) 2007 to 2009 Traffic Volume Counts Crash Frequency 2007 2008 2009 Correlation N Correlation N Correlation N Motorcycles Unweighted 0.156*** 40 0.853* 32 0.641* 21 Weighted 0.425* 0.914* 0.797* All Unweighted 0.174** 193 0.146** 197 0.058*** 198 Weighted 0.622* 0.626* 0.566* (b) 2010 to 2011 and average Traffic Volume Counts Crash Frequency 2010 2011 Average Correlation N Correlation N Correlation Motorcycles Unweighted 0.276*** 11 -- -- 0.481 Weighted 0.497*** -- 0.658 All Unweighted 0.137** 197 0.068*** 206 0.117 Weighted 0.602* 0.535* 0.590 *p < 0.01; **p < 0.05; ***p > 0.05 Table 13. Correlation of motorcycle crash frequency and the traffic volume at the nearest count station to the crash (Wisconsin data). The research team then added data obtained from Texas, Montana, and Wisconsin and found similar results. However, the Wisconsin data using straight-line distances were not as useful as using crashes on the same roadway as the count site and measuring the distances along that roadway. Two other findings are of interest. First, it is clear that the spatial distribution of motorcycle crashes is associated with the spatial distribution of traffic (and vice versa) to the point that a state can be confident in using crash location as an indica- tor of where (geographically) it should invest first in improved motorcycle count setups. Second, the logical extension is that

34 vehicles at those count stations may form the basis of a sur- rogate measure of motorcycle traffic volume. Because this association is a moderate level (with correlation values at about 0.50), such a surrogate measure would not have high precision; however, it would allow a state to quickly develop an estimate of the expected minimum and maximum values of overall (statewide) motorcycle traffic volume. The goal of this part of the project was to determine if motorcycle traffic (or total traffic) and crash frequency are spatially distributed in similar ways. If they are, it is reason- able to expect that states can avoid the expense of develop- ing separate count programs for motorcycles versus all other vehicles. The count stations can be sited in similar spots for all vehicles, including motorcycles. This is indeed what the results of this analysis indicate. Crash Prediction Model The research team conducted the crash data analysis while attempting to understand how to locate motorcycle count sites. The model described in Equation (4) has motorcycle crashes as the dependent variable, and the total traffic volume and motorcycle volume as the explanatory variables. Given that many non-flow related factors are known to affect the frequency of crashes, this model likely suffers from an omit- ted variables bias. However, the empirical assessment carried the methodology works equally well for weekends and week- days. That is, the locations of weekend motorcycle crashes can be used to determine where to conduct weekend counts just as the location of weekday crashes can be used to deter- mine where to conduct weekday counts. A further implication of these results is worth exploring although unproven in this research. The correlation between crash counts (weighted by the inverse distance of crashes from the nearest count station) and the total counts of all (a) 2007 to 2009 Traffic Volume Counts Crash Frequency 2007 2008 2009 Correlation N Correlation N Correlation N Weekday Unweighted 0.248* 191 0.245* 192 0.108*** 184 Weighted 0.605* 0.658* 0.566* Weekend Unweighted 0.032*** 180 -0.034*** 188 -0.054*** 185 Weighted 0.513* 0.442* 0.465* (b) 2010 to 2011 and average Traffic Volume Counts Crash Frequency 2010 2011 Average Correlation N Correlation N Correlation Weekday Unweighted 0.208* 190 0.114*** 196 0.185 Weighted 0.636* 0.551* 0.603 Weekend Unweighted -0.027*** 175 -0.043*** 181 -0.025 Weighted 0.372* 0.432* 0.445 *p < 0.01; **p < 0.05; ***p > 0.05 – 0.1 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 Unweighted Weighted Pe ar so n' s R Correlation Category Weekday Weekend Figure 16. Weekday and weekend crash frequency (weighted and unweighted) correlated with motorcycle and total vehicle counts at the nearest count station (Wisconsin data). Table 14. Weekday and weekend crash frequency (weighted and unweighted) correlated with motorcycle and total vehicle counts at the nearest count station (Wisconsin data).

35 model development considered each year separately. Table 15 summarizes the results of the negative binomial (NB) model. The Pearson c2 statistic for the model is 113.8, and the degrees of freedom are 107 (= n - p = 111.4). As this statistic is less than c20.05,107 (= 132.1), the hypothesis that the model fits the data cannot be rejected. The coefficient for total traffic volume is below 1.0, which indicates that motorcycle crash risk increases at a decreasing rate as traffic flow (all vehicles) increases. It should be noted that the 95% confidence intervals for each of the coefficients did not include the origin, meaning all the coefficients in the model are significant at the 5% level. Figure 17 shows the estimated crashes with the change in motorcycle volume when the total traffic flow changes from 20,000 vehicles per day to 100,000 vehicles per day. This fig- ure illustrates that the motorcycle crash risk increases at an increasing rate as the motorcycle volume increases. out in this work will still provide valuable insights and poten- tial applicability because the model can be easily recalibrated and applied to other states. SegLength TotVol MCVolMC Crashes e e (4)10000 1 2= × × ×β β β × where: SegLength = Length of segment in the vicinity of count site (5 miles either side of the count site is considered). TotVol = Total volume at count site. b1, b2, b3 = Model coefficients derived from state database. MCVol = Motorcycle volume at count site. The crash prediction model relied on the Wisconsin data- set and used crash data collected from 2007 to 2011. Because of the large variability in traffic flows between different years, 0 5 10 15 20 25 0 500 1000 1500 2000 2500 M ot or cy cl e Cr as he s/ ye ar /m ile Average Daily Motorcycle Traffic Total Vol=20,000 Total Vol=50,000 Total Vol=100,000 Variable Estimate Standard Deviation t-statistic Constant )( 0β -8.0134 1.064 -7.53 TotVol )( 1β 0.8050 0.117 6.88 MCVol )( 2β 0.7671 0.391 1.96 Dispersion parameter (α ) 0.6262 0.112 5.58 Log-likelihood -320.5 Akaike information criterion AIC 649.1 Pearson 2χ 113.8 ( 1.1322 107,05.0 =χ ) Sample size (N) 111 † Standard error Table 15. Estimates of the negative binomial model. Figure 17. Relationship between motorcycle traffic volume and motorcycle crashes.

36 sify vehicles based on vehicle length alone. Inductive loops (if used alone) and magnetometers use magnetic length as an estimate of physical length. The magnetic length of motor- cycles is usually less than the physical length and often is even less than the wheelbase. For each of the five test systems discussed, tabulated results are included. In the fourth column of each table is a ratio that shows the number of motorcycles detected by that system divided by the ground truth count. (Each table also indicates what technology was used to establish the ground truth). In each case, the ratio in the fourth column is converted to a percent to establish the simple detection accuracy. The last column to the right indicates that the available data were pro- vided as per vehicle (PV) records as needed. Hourly bins were used depending on output options available from each system. Infrared Classifier Testing of this system involved two motorcycle rally sites: one in New Ulm, Texas, on May 18, 2012, and the other in Daytona Beach, Florida, on October 20, 2012. The New Ulm site used a two-lane roadway with width of about 18 ft with no shoulders or pavement striping so vehicles often trav- eled in the center of the road instead of on the right side. The speeds on this road were between 30 and 45 mph and the traffic volume was very low. The US-92 test site used for the Florida data collection was a four-lane divided highway with depressed grass median and speeds ranging from 50 mph to 60 mph. Data collection only involved the two eastbound lanes, and the traffic was always free-flowing with no interruptions. The overall traffic volume was relatively high, and the number of motorcycles passing the test site was high. Table 16 summarizes the results of testing of the IR Clas- sifier. Notice the ratio in the fourth column—the number of motorcycles detected by the IR Classifier divided by the actual (ground truth) count. For example, in Table 16, for May 18, 2012, the detected motorcycle count of 163 is the number clas- sified by the IR system as motorcycles or Class 1 vehicles, and the actual (ground truth) count is 134. Converted to a deci- Field Data Findings This section presents results for each of the five selected detection systems, one by one, followed by a summary. The discussion addresses the following key components to help properly interpret the results: • Description of test conditions (e.g., road geometry, traffic volume, speed) • Performance in detecting motorcycles • Initial cost of the system • Whether portable or fixed • Technology advantages and disadvantages • Appropriate applications of the technology Under the category of technology advantages and dis- advantages, one aspect that might not be well understood is how each detector determines vehicle classification. Detec- tors normally do this in one of two ways: either they detect a length or they detect axles and spacing between axles to determine the vehicle classification. The exception is the multi-technology detector, which relies on three technologies (acoustic, IR, and a stereo camera) to determine a vehicle class. The measurement using axles and spacings typically leads to a classification using FHWA Scheme F, which has 13 vehicle classes. Of the technologies tested in this research, only the inductive loop/piezoelectric sensor systems and the IR Clas- sifier have the ability to classify directly according to FHWA Scheme F. For vehicle classification reporting purposes, the FHWA allows some states to collect data according to length but requires results to be reported as Scheme F classes. Large motorcycles (Class 1 vehicles in Scheme F) are generally less than 7 ft in length (unless they have been modified) and the smallest Class 2 vehicles start at about 7 ft in length. Detectors that measure only vehicle length can do so based on either a physical length estimate or a magnetic length estimate, but states that use these systems must be able to create length bins that tie directly to one or more vehicle classes within Scheme F. Of the five technologies tested in this research, magnetometers and the tracking video clas- Date and Location Time Span Ground Truth Motorcycles Detected/Actual Simple Detection Accuracy Bin May 18, 2012 Texas 13:00–18:46 Video 163/134 121.64% PVa Oct. 20, 2012 Florida 07:30–09:30 Video 709/744 95.30% PV a PV = per vehicle Table 16. Data collection summary for the IR classifier.

37 persons. With two persons, one person was stationed on each side of the monitored roadway to make adjustments to the aim of the sender-receiver system via an iterative process. The exception was the October 20 setup in Florida, where only one person was available to perform the setup. Both the transmitting and receiving components of the IR Classifier system have built-in sights to assist in the aiming procedure. The lead person also has access to software on a laptop that guides the setup to optimize the aim of each com- ponent. Even then, monitoring the passage of several vehicles followed by additional adjustments appears to be required to get the best results. Setup required about 30 minutes on both October 19 and October 20, but the Texas setups took lon- ger. Overall, the setup appears to be fairly complex, especially with portable units like the ones used for this research; how- ever, a 30-minute setup time is certainly acceptable. Setup of a permanent site (for long-term data collection) requires a site survey and permanent enclosures on both sides of the roadway to ensure initial and continuous alignment of the beams. The manufacturer was developing an improved user’s guide to assist in setup of the system, but the guide was not available as of February 2013. Strengths of the IR Classifier are as follows: • It can classify according to FHWA Scheme F (13 classes). • Its accuracy for motorcycle detection is good. • It is non-intrusive and highly portable. • The portable unit is self-contained with its own power supply, and newer fixed units are Ethernet-equipped for direct firmware upgrades and other communication needs. mal, the simple detection accuracy on May 18 was 121.64%. The available data were provided as PV records as needed. The IR Classifier motorcycle detection results at the Texas rally are clearly not as good as those at the Florida rally. Setup difficulties in Texas were likely a factor in these results. A criti- cal factor guiding site selection in both Texas and Florida was that neither the Texas DOT nor the Florida DOT allowed placement of the IR Classifier units near high speed travel lanes without guardrail protection. Also important to site selection were the pavement cross-slope and crown. These factors are critical to getting good results given that the IR beams cross the roadway underneath the vehicle and must not be interrupted by anything but vehicle tires. Closer scrutiny of the IR Classifier results indicated 34 false alarms at the Texas rally compared to only 5 false alarms in Florida in a much larger dataset, leading to a relatively large number of overcounts in Texas. The difference might be due to site-specific conditions, such as very slow speeds at the Texas rally on May 18 or differences in the setup. The false alarms at the Texas rally were due largely to the presence of Class 9 vehicles (5-axle tractor-semitrailers) and the incorrect clas- sification of tandem axle groups as motorcycles. However, the Florida dataset included several Class 9 vehicles as well, and they were all classified correctly. Based on conversations with early users, the IR Classifier was not as accurate during certain inclement weather condi- tions, such as heavy rain. Based on recent comparisons dur- ing a rainy season in Australia, the effects of rain seem to have been mitigated; however, the Australian results need to be verified by local tests that might include a variety of weather conditions besides just rain. The manufacturer of the IR Classifier provided cost information based on the Indiana DOT contract prices. The cost is $23,190 for the system (transmitter/receiver, software, laser site, and power and communication cables). Portable cabinets (two each) for enclosing the electronics at the roadside cost $3,660 per set. Included in the cost are the two cabinets, a solar panel, a solar regulator, leveling legs/ feet, rechargeable harness, locking lid, and battery com- partment for two 12V 7Ah batteries. Figure 18 is a photo of the cabinet offered by the IR Classifier distributor. The distributor does not sell the portable tripod setup that was used for this research but uses it strictly for demonstration purposes. The total cost for the complete package, including the transmitter/ receiver with its components and the two portable cabinets would be $26,850. This system has sufficient power to collect data for 48 hours. An external battery pack (additional cost of $1,050) can extend the data collection time to 7 days. The distributor insisted that only trained personnel be used to set up the unit at all sites for this project, so researchers were unable to obtain a good understanding of how difficult or easy setup might be. All setups except one used two trained Source: Control Specialists, used by permission. Figure 18. Photo of IR classifier cabinet with solar panel.

38 This roadway had an AADT of about 50,000 vehicles per day and had a speed limit of 70 mph. Texas DOT reinstalled this system in an LPL configuration in January 2012 at the same time it reinstalled the ADR-6000 inductive loops following a new hot mix asphalt overlay. The inductive loops and piezo- electric sensors are connected to an IRD TRS Rack II clas- sifier, which is normally programmed to collect data in 15 vehicle classes and store them in 1-hour bins. Modifying the data collection parameters was a source of problems, so some of the test results are reported as hourly summaries and oth- ers are PV results. Table 17 summarizes the data collection results for the inductive loop/piezoelectric sensor system. Appendix B shows a data sample to indicate the original format. Results indicate a simple detection accuracy ranging from a low of 0 percent to as high as almost 90 percent. After research- ers discovered the low detection accuracy of this system and alerted Texas DOT, an investigation into the issue began. Texas DOT could not correct the problem in time to redo the data collection and perhaps achieve better results. According to Texas DOT, the initial cost of installing a four-lane inductive loop/piezoelectric sensor system like the one at College Station would be about $61,000, but cost would vary depending on the length of road bores, number of ground boxes, and length of conduit runs. Components included in the referenced cost are the controller, electronics, modem, and an 80W solar panel. Advantages of full lane-width piezoelectric sensors with inductive loops include: • The system is able to collect 13 vehicle classes consistent with FHWA Scheme F. • The technology is mature. • Necessary components are widely available. • The cost of a portable system like the one used on this proj- ect ($26,850) is acceptable given than it is reported to cover as many as four or more lanes. • Its user interface, while likely complicated for a new user, is obviously adequate for expert users to get acceptable results. Weaknesses of the IR Classifier are as follows: • Site selection for portable TIRTL installation is critical for getting good results. • Sites for permanent installations require an extensive site survey and likely additional cost. • Battery power (without the add-on batteries) is limited to 48 hours. • Setup by a novice user will likely be difficult and require a lengthy learning period. • Vandalism could be an issue because the portable units are at ground level and vulnerable. The portable IR Classifier appears to have a wide range of applications, but site selection is critical to achieving good results, especially for portable units. The distributor has stated that the Illinois DOT is using one permanently installed IR Classifier system to cover eight lanes on a high volume freeway and is satisfied with the results. The research team has not verified this claim with the DOT. The configura- tion tested in this research has excellent portability; it can be dismantled in a matter of minutes and placed in a compact box for transfer to the next site. Inductive Loops/Piezoelectric Sensors Testing of the inductive loop/piezoelectric sensor system occurred at one site, the SH 6 test facility in College Station. Date Time Span Ground Truth Motorcycles Detected/Actual Simple Detection Accuracy Bin June 30, 2012 11:00–12:00 09:00–10:00 Video/ ADR-6000 5/6 2/3 Hourly July 1, 2012 11:00–12:00 09:00–10:00 Video/ ADR-6000 0/3 0/4 Hourly July 3, 2012 11:00–12:00 11:00–12:00 Video/ ADR-6000 4/24 10/20 Hourly July 21, 2012 00:00–24:00 ADR-6000 104/191 Hourly July 22, 2012 00:00–24:00 ADR-6000 76/154 Hourly July 23, 2012 00:00–24:00 ADR-6000 41/73 Hourly Feb. 8, 2013 13:00–15:00 Video/ADR 20/102 0% 0% 88.3% 66.7% 16.7% 50.0% 54.45% 49.35% 56.16% 21.05% PV a a PV = per vehicle Table 17. Data collection summary for inductive loop/piezoelectric sensor system.

39 of 70 mph. A manufacturer’s representative reinstalled this system in January 2012, while Texas DOT had the lanes closed for reinstallation of the ADR-6000 inductive loops and the inductive loop/piezoelectric sensor system following a pave- ment resurfacing project. Table 18 summarizes the results of magnetometer testing. The ground truth for these tests was established using a com- bination of recorded video and the Peek ADR-6000 classifier. Magnetometer components include the four sensor nodes in all lanes, an Access Point (AP) for communication and data storage, and electronics in the cabinet. Communication from each of the sensor nodes is wireless to the AP at the roadside, with node power provided by an internal battery. The most common usage involves simple on and off signals (vehicle present or not) much like inductive loops. The research team downloaded data for test purposes via the Internet by access- ing the Sensys Networks server in Berkeley, California. Appen- dix C shows the original data format. The vehicle length in ft is the key value for matching these data with ground truth data. This sample is sorted by length (short to long), indicat- ing a number of short vehicles as estimated by each vehicle’s magnetic length. The manufacturer does not market the wireless magne- tometer system as a motorcycle detection system, so this research is the first known effort to quantify the accuracy of these detectors for motorcycles. The positioning of the sen- sor nodes at SH 6 was intended to accommodate the widest variety of uses for current and future research purposes, not just for this project. This research kept all five nodes turned on in each lane for all tests; however, only motorcycles travel- ing in the center of the lanes generated a speed and vehicle length (i.e., due to the 1-3-1 pattern, motorcycles centered in the lane crossed three center sensor nodes). Because this was the first motorcycle test for this detec- tor, its detection attributes were not well understood at first, either by the researchers or by the manufacturer’s representa- tive. For that reason, perhaps, and also because there might be unwanted adjacent lane detections with higher sensitivity, the manufacturer was hesitant to adjust the sensitivity to improve • How to install the components is common knowledge. • The system is immune to weather and light conditions. Disadvantages of inductive loops/piezoelectric sensors include: • Sawcutting the pavement compromises pavement integrity. • Piezoelectric sensors can fail prematurely and unexpect- edly, sometimes with no warning. • The installation places workers in close proximity to traffic. • Installation and maintenance can cause traffic delays. • Underground components are sometimes damaged by other roadside construction. • Once installed, the system is not as flexible as some other systems. • The life-cycle cost of loops and piezos could be higher than that of competing systems (e.g., if installed in weak pavement). Applications of inductive loops and piezoelectric sensors are fairly widespread on free-flow roadways in both rural and urban areas, although many jurisdictions still use 6 ft piezos instead of those that cover the full lane-width. The induc- tive loop/piezoelectric sensor results from this research per- taining to motorcycles would likely be worse with 6 ft piezos. Loops alone are still used by many agencies for traffic signal detection and on freeways for collecting data such as vehicle speeds, counts, and occupancies. Given the disadvantages noted for this technology and the availability of other detec- tion systems today, many agencies are abandoning inductive loops, at least as they fail, and are replacing them with less- intrusive devices. Magnetometers Testing of the wireless magnetometer system involved one site, the SH 6 test facility at College Station. SH 6 has an AADT of about 50,000 vehicles per day and has a speed limit Date Time Span Ground Truth Motorcycles Detected/Actual Simple Detection Accuracy Bin July 1, 2012 10:00–11:00 ADR-6000 2/3 66.67% Hourly Feb 8, 2013 13:30–15:00 Recorded video 76/97 78.35% PV a Feb 22, 2013 15:00–16:00 Recorded video 11/16 68.75% PV a PV = per vehicle Table 18. Data collection summary for magnetometers.

40 data to a client server, but another option is to store data on a data server in Berkeley for an additional cost. The manufac- turer now provides a 5-year warranty for major components of its system. The data results from the tests at the TTI facility on SH 6 indicate that the sensor configuration was not optimized for motorcycle detection and could probably be improved. This magnetometer system has the following positive attributes: • It takes less time to install than inductive loops and is less damaging to pavements. • Its accuracy is similar to inductive loops, especially for larger vehicles. • It uses wireless communication, so damage by other road- side work is minimized. • The cost to cover four lanes is $15,964, and for two lanes, $10,204 (with 6 sensors per lane). Disadvantages of magnetometers include: • The system requires special configuration to detect motor- cycles accurately. • It is placed in the pavement, so traffic control is required. • Pavement milling and resurfacing would destroy the sen- sor nodes (although they can be removed in advance of the milling operation and reused). Applications of wireless magnetometers have become widespread in both urban and rural areas. Early concerns with battery life have been addressed, and the AP has been modified recently to be much more amenable to solar power. Magnetometers are used by many agencies for traffic signal detection and on freeways for collecting data such as vehicle speeds and counts. Multi-Technology System Testing of this system involved two sites—the SH 6 test facility and the motorcycle rally in New Ulm, Texas. Table 19 indicates a few of the count periods used to test this detector. Following the rally, the manufacturer modified the detector and later returned it to the researchers for further tests. In its initial form, the detector stored an image of the detected vehi- cle, and the storage process apparently overwhelmed the pro- cessor at high speeds. Following the September 2012 tests at SH 6, the manufacturer continued to improve the detector for future tests. For this research, the simple detection accuracy for several days and the overall detection accuracy for May 19, 2012, at the Texas rally indicate that it needs improvement. Its performance was modestly better at the rally, where speeds were slower—in the range of 30 mph to 50 mph. Considering individual hourly results for September 21, the best 1-hour motorcycle detection. As data collection and evaluation con- tinued, analysts hypothesized that motorcycles needed to pass closer to the sensor nodes than originally thought (i.e., to within about 1.0 ft) to achieve consistent and predict- able results. Close scrutiny of the February 8, 2013, data for a test in which selected riders were instructed to ride in the wheelpaths indicated that the outside sensor nodes detected motorcycles almost every time, but the center nodes did not. Randomly passing motorcyclists usually pass in the wheel- paths as well, leading to what appeared to be a high percent- age of missed detections. This finding led to another test on February 22, 2013, dur- ing which selected riders were instructed to travel over the nodes in the center of the lane. This dataset indicated that the magnetometers detected motorcycles about 70 percent of the time when passing through the established detection zone. Some riders might have missed the tiny (2 ft-wide) detection zones, and the detection rate could be even higher with better understanding of how to establish the sensor spacings and sensitivities. Like most other systems, the magnetometer system mea- sures vehicle length by calculating a speed, then determining vehicle length based on the speed and presence time. To do this requires two stations positioned a known distance apart. Detection of each vehicle occurred with single magnetom- eters on each end of the 1-3-1 pattern. A better configuration for motorcycles would be two (or three) nodes at each of the two endpoints to ensure improved detection in the wheel- paths, with the endpoints at least 12 ft apart. The research team has found that this system’s accuracy for larger vehicles is similar to that of inductive loops. Portability of this system could be improved by using a surface-mount sensor node, but the manufacturer currently has no known plans to market such a sensor. A few years ago, the manufacturer provided a few prototype surface-mount sensors for research purposes, but the manufacturer did not want them used where motorcycles might travel. Their con- cern was that motorcyclists could lose control if their tires struck the raised sensors. Batteries in the depressed sensors last about 10 years, so an operating agency could install the depressed sensors as part of a semi-mobile system in which other components are moved from site to site. The agency could mount the other components on a mobile platform such as a trailer with a telescoping pole and a solar panel for power. This sce- nario could use the mobile components for as many sites as needed while placing the fixed sensor nodes to be used only when data collection is needed. The cost of six sensor nodes per lane (at $480 each), epoxy, extension card in the cabi- net, the AP with Ethernet ($3,400), and other components (excluding the trailer) would total $15,964 for a four-lane system and $10,204 for a two-lane system. The most com- mon option for data storage is for the manufacturer to push

41 • Its user interface needs further development. • In its current configuration, it can only cover one or two lanes. The multi-technology sensor has considerable potential for future data collection in situations for which a low-cost, non-intrusive sensor is needed. However, it is not ready for widespread use at this time. The manufacturer has secured additional development funding through the SBIR program. The technology appears to be conducive to solar/battery power, and its communication needs can be met with a low- bandwidth solution. Tracking Video System Testing of the tracking video detector involved two sites— the SH 6 test facility and the motorcycle rally in New Ulm, Texas. Table 20 shows representative results of field testing. The mounting system for the rally was a van owned by Clemson University that was driven to the site for this purpose. Its tele- scoping pole was about 35 ft high and its offset from the road at both data collection sites at New Ulm was about 15 ft at the May 18 site and about 25 ft at the May 19 site. However, there were no occlusion issues at the rally because of the low traffic volume. Tests in College Station involved mounting the FLIR® detection rate for the Migma system was 77.78 percent (cor- rectly detected 7 of 9 motorcycles). None of the tests of the multi-technology sensor involved inclement weather. For this detector, the research team would anticipate a decline in performance during heavy rain or fog. Motorcyclists do not normally ride during heavy rain, so this might not be a significant factor. As noted before, however, traffic speeds were a challenge, as were traffic volumes with the earlier version of the detector. The setup of this detector was probably more difficult because of its newness and com- plicated user interface. Advantages of the multi-technology detector are: • It has a low initial cost (estimated by the manufacturer to be about $6,000). • It is non-intrusive—it can be mounted on a pole beside the roadway. • It has a compact size. Disadvantages of the multi-technology detector are: • Its accuracy for motorcycle detection must be improved. • It is not designed to detect anything but motorcycles, so a user would have to install a second device to detect non-motorcycles. Date Time Span Ground Truth Motorcycles Detected/Actual Simple Detection Accuracy Bin May 19, 2012 09:00–12:00 Video 143/206 69.42% PV Sept. 5, 2012 09:20–10:30 ADR-6000 26/45 57.80% PV Sept. 21, 2012 17:00–22:00 ADR-6000 21/46 45.65% PV Sept. 22, 2012 17:00–20:00 Video 13/22 59.09% PV Sept 23, 2012 17:00–20:00 Video 6/21 28.57% PV a PV = per vehicle Table 19. Data collection summary for the multi-technology detector. Date Time Span Ground Truth Motorcycles Detected/Actual Simple Detection Accuracy Bin May 18 (day) 15:00–20:40 Video 111/168 66.07% PV a May 18 (night) 20:40–21:00 Video 9/12 75.00% PV May 19, 2012 09:00–12:00 Video 211/236 89.41% PV June 30, 2012 10:00–12:00 Video 14/18 77.78% PV July 1, 2012 11:00–12:00 Video 2/3 66.67% PV July 3, 2012 09:00–12:00 Video 46/50 92.00% PV a PV = per vehicle Table 20. Data collection summary for tracking video.

42 the DVR is needed only if the agency wants to record video as opposed to recording classification counts. The laptop can store a maximum of about 1 hour of video internally without a DVR. An agency might want to store some video for verification purposes while simultaneously storing clas- sification counts. Real-time processing in the field would again involve two costs—one for the initial camera cost ($500 to $1,500, or higher for IR cameras) and the other for camera power (assuming no data communication). The laptop battery life should be sufficient for several days of data collection depending on the amount of processing, but it would require an auxiliary power source for longer periods. Communica- tion methods for the laptop can vary depending on the net- work architecture. It can output data to agency personnel on-site in real time as well as over fiber or wireless networks. The unit comes with a mobile broadband card built into the laptop but would involve user fees. The setup of the tracking video system was reasonably easy, and the system is reasonably portable. Moving it from one site to another could probably be accomplished by using a trailer and telescoping pole provided the pole could be stabi- lized to minimize camera movement during data collection. Setup for the tracking video system was easier than that for most other video systems encountered by the research team and only required setting markers (e.g., traffic cones) at pre- determined locations within the viewing area. The tracking video detector’s strengths include: • It is a non-intrusive system. • Video tracking is considered more accurate than tripwire video systems for difficult vehicles like motorcycles. • The basic cost for the video system to cover four lanes (or possibly six lanes, depending on the layout) would require one laptop unit ($12,000), IR camera ($2,500), plus ancil- lary cables and mounting equipment (about $500), for a total of about $15,000. • As with any video detector, this system provides a view of the roadway that can be used for verification purposes. The tracking video detector’s weaknesses include: • Its performance is likely to be affected by some weather con- ditions, such as heavy rain and fog (although this research did not confirm this assumption). • Camera mounting situations may occur for which glare and day/night transitions cause diminished performance (a flaw that is common with most video detection systems). • It would need to use existing stationary poles, or it could be mounted on a trailer with a telescoping pole. • Vehicle classification is based on length, with five classifica- tion bins available. camera on a 5-ft riser supported by the lower pole mast arm. This mounting location placed the camera height at about 25 ft with an offset from the nearest (southbound) lane of about 20 ft. Tests in College Station used only the south- bound lanes. Most of the testing occurred during daylight, although the period on May 18 at the New Ulm rally from 8:40 p.m. until 10:00 p.m. was after dark. Statistical tests using the IR camera comparing day versus night detections revealed that day detections were biased toward false alarms whereas night detections were biased toward misses. Nighttime errors were more prevalent with the standard camera, so the nighttime analysis only used the IR camera results. All tests at SH 6 used the standard camera because of a wavy image from the IR camera. Researchers attempted to troubleshoot the problem but were unable to find a solution in a timely manner and were under pressure to return the FLIR camera. The simple detection result indicates that nighttime detection using the IR camera was not too different from the daytime, although the sample size was small. In all cases of night detection using recorded video for ground truth either at this rally or in Col- lege Station, there was insufficient lighting to accurately determine vehicle type with a high degree of certainty. There are two use-cases for the tracking video technology in terms of how data might be collected/processed: • Recorded video post-processing (generally off-site) • Real-time processing in the field where data storage occurs on-site Overall, the cost to the consumer regarding the technology would be the same for both scenarios, but the ancillary equip- ment, such as the camera and the power and communication needs, would differ. For both scenarios, the user would have a field-hardened laptop PC with software loaded. The sys- tem used in this research retails for $12,000 and includes two video streams simultaneously being processed by the laptop and 1 year of hardware warranty and software add-ons and bug-fixes at no additional charge. Users also have the option of adding third and fourth “channels,” at an additional cost of $3,000 each, to allow simultaneous processing of three or four video streams. The unit costs are one-time fees, and there are no monthly or hourly usage, data processing, or user license charges. In-office post-processing would involve two costs—one for the initial camera cost ($500 to $1,500, or higher for IR cameras) and the other for camera power and possible cost for data communication. IR (or thermal) cameras cost at least $2,500. In lieu of communication costs, an agency might choose to record its own video stream using a digi- tal video recorder (DVR). The DVR cost might range from $1,000 to $3,000, depending on available features. However,

43 Classification of Non-Motorcycles All of the selected detectors except the multi-technology detector have the ability to classify vehicles besides motor- cycles. However, only two can classify vehicles based on axle detection using all 13 classes contained in FHWA Scheme F: the inductive loop/piezoelectric sensor system and the IR Classifier. Given that motorcycles typically are only a small proportion of total vehicles, when choosing a detector it is important that decision-makers know the classification accu- racy for vehicles other than motorcycles in order to make the best overall choice. Applications of this technology include both rural and urban count sites where adequate poles or other supports already exist. Purchasing a new pole at each site would increase the overall cost substantially. The system could potentially be trailer-mounted for monitoring remote areas. In that case, the trailer would need to provide power to the laptop, the camera, and the DVR, potentially via solar/battery power. Summary of Field Data Findings Table 21 summarizes the results of all systems tested that were provided individually above. Date Time Span Ground Truth Motorcycles Detected/Actual Simple Detection Accuracy Bin Infrared (IR) Classifier May 18, 2012 13:00–18:46 Video 129/134 96.27% PV Oct. 20, 2012 07:30–09:30 Video 709/744 95.30% PV Inductive Loop/Piezoelectric Sensor System June 30, 2012 11:00–12:00 09:00–10:00 Video/ ADR-6000 5/6 2/3 88.3% 66.7% Hourly July 1, 2012 11:00–12:00 09:00–10:00 Video/ ADR-6000 0/3 0/4 0% 0% Hourly July 3, 2012 11:00–12:00 11:00–12:00 Video/ ADR-6000 4/24 10/20 16.7% 50.0% Hourly July 21, 2012 00:00–24:00 ADR-6000 104/191 54.45% Hourly July 22, 2012 00:00–24:00 ADR-6000 76/154 49.35% Hourly July 23, 2012 00:00–24:00 ADR-6000 41/73 56.16% Hourly Feb. 8, 2013 13:00–15:00 Video/ADR 20/102 21.05% PV Magnetometers July 1, 2012 10:00–11:00 ADR-6000 2/3 66.67% Hourly Feb 8, 2013 13:30–15:00 Rec. Video 76/97 78.35% PV Feb 22, 2013 15:00–16:00 Rec. Video 11/18 61.11% PV Multi-technology Detector May 19, 2012 09:00–12:00 Video 143/206 69.42% PV Sept. 5, 2012 09:20–10:30 ADR-6000 26/45 57.80% PV Sept. 21, 2012 17:00–22:00 ADR-6000 21/46 45.65% PV Sept. 22, 2012 17:00–20:00 Video 13/22 59.09% PV Sept 23, 2012 17:00–20:00 Video 6/21 28.57% PV Tracking Video Detector May 18 (day) 15:00–20:40 Video 111/168 66.07% PV May 18 (night) 20:40–21:00 Video 9/12 75.00% May 19, 2012 09:00–12:00 Video 211/236 89.41% PV June 30, 2012 10:00–12:00 Video 14/18 77.78% PV July 1, 2012 11:00–12:00 Video 2/3 66.67% PV July 3, 2012 09:00–12:00 Video 46/50 92.00% PV a PV = per vehicle a Table 21. Summary of results for all five detectors.

44 Magnetometers The research team compared the magnetometer hourly data with classification counts from the Peek ADR-6000. Table 23 summarizes hourly totals summed to form a daily (24 hr) comparison. These results indicate that daily sums are very close to the ground truth values, differing by no more than 2 to 3 percent. Appendix C indicates the original data output format from the magnetometers. Multi-Technology System The multi-technology system currently classifies only motorcycles. Tracking Video System The tracking video system classifies vehicles according to the following length or vehicle bins: • Motorcycles • Cars and pickups • Single-unit trucks and buses Infrared Classifier Analysis of non-motorcycle detection by the IR Classifier did not reveal any particular weaknesses for any vehicle class. The analysis used about an hour of the Florida data collected on October 20, 2012. Within this period, the IR Classifier detected a total of 798 Class 2 through Class 9 vehicles. Based on manual observation of the recorded video used to estab- lish ground truth for the same period, the IR Classifier cor- rectly classified 783 of these vehicles for a simple detection accuracy of 98.12 percent. Inductive Loops/Piezoelectric Sensors Table 22 summarizes the hourly totals for non-motorcycles from the inductive loop/full lane-width piezoelectric sensor system installed by Texas DOT in January 2012. This analy- sis combines FHWA classes 2 through 13 into three groups for simplicity. Results are better for larger vehicles than for Class 2 and Class 3, but the analysis reveals problems with the system part way through the data analysis. Properly operating inductive loop/piezoelectric sensor systems typically perform much better. Vehicle Class* Time Non-Motorcycles Detected Loop/Piezo Accuracy Video Loop/Piezo Group 2 (Class 2 and Class3) 10:00–11:00 24 4 16.67% 11:00–12:00 20 10 50.00% Group 3 (Class 4 and Class 5) 10:00–11:00 1036 1074 103.67% 11:00–12:00 1186 1232 103.88% Group 4 (Class 6 to Class 13) 10:00–11:00 83 90 108.43% 11:00–12:00 1186 1232 103.88% *Group 1 (Class 1) is not included in this table. Date Lane 1a Lane 2 Lane 3 Lane 4 April 17, 2011 -1.08% -1.82% -0.94% -0.19% April 18, 2011 -1.06% -1.99% -1.28% 0.01% April 19, 2011 -1.00% -2.03% -1.18% -0.08% April 20, 2011 -1.01% -2.63% -1.12% -0.03% April 21, 2011 -1.97% -2.45% -1.60% 0.05% a Lane 1: northbound right lane, Lane 2: northbound left lane, Lane 3: southbound left lane, Lane 4: southbound right lane. Table 22. Classification of non-motorcycles by inductive loops/ piezoelectric sensors. Table 23. Classification of total traffic by the magnetometers.

45 • Combination trucks • Other/unknown Using the dataset for July 3, 2012 (daylight only and good weather conditions), analysts found that the tracking video detection accuracy was best for the car/pickup bin but over- counted single-unit trucks/buses and undercounted combi- nation trucks. Observations of the video indicated that most of the errors were due to either occlusion or not being able to distinguish between two vehicles in close proximity. Table 24 summarizes these results. Vehicle Type Number by TrafficVisionTM Ground Truth (Video) Percent Correct Cars and pickups 3,336 3,396 98.23% Single-unit trucks/buses 201 159 126.42% Combination trucks 170 218 77.98% Table 24. Classification of non-motorcycles by the tracking video system.