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.
53 C h a p t e r 5 General Methodology One of the goals for this project was to explore the utility and usability of existing data in traffic safety analysis, specifically in the investigation of surrogate measures of safety. The approach requires vehicle trajectories. The data collected from the Cooperative Intersection Collision Avoidance Sys- tems (CICAS) project intersections, at least at first glance, looked like a promising source of information. This infra- structure was designed for the purposes of the CICAS project, however, and certain features constrain the usefulness of these data for more general research purposes. In adherence to the principal objective of the SHRP 2 Safety Project S01, effort was devoted to assessing the potential utility of this extensive data source. This was done both to test the projectâs core methodology and to identify possible future improve- ments to the data-collection infrastructure. This chapter summarizes the findings of the analysis and the problems encountered, along with their proposed solutions, and con- cludes with an exploratory analysis of one near-crash event. During the course of this project, all CICAS project sites were explored as potential data sources. The Minnesota Highway 52 site, although it has more recorded crash cases, deploys older technology; the North Carolina US-74E site has better sensors and corrects some errors in the collection and postprocessing of data. These were the main reasons for focusing the projectâs efforts on the data originating from the North Carolina site. The site is illustrated in Figure 5.1; US-74 is the main road and Strawberry Boulevard is the side road. The goal was to attempt to isolate near-crash events and use the structural-model methodology to investigate the background and input conditions characterizing a crash or near-crash event and their relation to the occurrence or nonoccurrence of a collision. The data harvesting procedure involved the extraction of possible cases from the CICAS database, use of filters for the removal of false positives, and visual inspection of the resulting trajectories. It is important to note that in contrast to the other data sources used in this project, the magnitude of the data available from the CICAS site made necessary the development of automated methods for event data extraction and the creation of utilities that accelerate manual inspection of the extracted information. The CICAS vehicle trajectory database structure and avail- able information is described in Appendix B. In summary, the information available for each vehicle includes x, y coor- dinates in the state plane system, speed, and acceleration measurements, all at 10 Hz. It is vital to understand that these are not the raw collected measurements from the site but are derived from the raw measurements provided by the instru- mentation during a data-reduction stage. From the detection hardware used, it is known that the sensor radar-based mea- surements for main-approach vehicles are range, range rate, and azimuth at 10 Hz, with similar information collected from the LIDAR sensors for the side-road vehicle (although it is unclear at what interval). The postprocessing of the raw data includes filtering as well as projections to known road- way features that introduce some difficulties for this study. For example, the position of the vehicle is always projected onto the lane centerline, resulting in some loss of informa- tion regarding the transition during lane changes, but more importantly during evasive actions taken by the drivers to avoid collisions. In addition, after closer inspection of the data, the project team concluded that the side-road vehicle position is not updated at 10 Hz, since the provided data gen- erate a stepwise trajectory profile. Speed and acceleration for the side-road vehicle display apparently unrealistic behavior, so they were excluded from this investigation. Figure 5.1 shows the North Carolina site for the CICAS project. The CICAS database codes events based on side-road vehicle maneuvers, along with the gap duration from the main-road vehicle movements. This database postprocessing makes possible the harvesting of completed cross-mainline maneuvers, where the accepted gaps of the side-road crossing vehicle may have been small. This involves finding a plausible Analyses Using CICAS Site-Based System
54 time envelope of small, accepted lag times during the cross- mainline maneuver over the complete lifetime of the tracked vehicle. The search was separated by direction for the two cross-main-road maneuvers (one for each direction). The technique is illustrated for one of the maneuvers. Essentially, it involves two inner join steps. The first step is the inner join of all tracked targets associated with small gap events for the given maneuver within the observation lifetime of the tracked target that completed the maneuver (i.e., when the target was first and last seen by the system): CREATE VIEW closecalls AS SELECT targetid, maneuver, veh_class, sub1_acpt_lag_zone_exit_date, sub1_acpt_lag_ zone_exit_time FROM vehicle_acpt_lag WHERE (maneuver = 1) AND ((3.0 > ANY(sub1_acpt_lags)) OR (3.0 > ANY(sub2_acpt_lags))); CREATE VIEW lifetime AS SELECT vt.targetid, vt.zone_ first_seen, vt.date_first_seen, vt.time_first_seen, vt.zone_ last_seen, vt.date_last_seen, vt.time_last_seen FROM vehicle_time AS vt INNER JOIN closecalls AS cc ON (vt.targetid = cc.targetid); The second step uses the lifetime of this target to harvest all tracked targets present during the evolution of the accepted lag times for the maneuver. SELECT * FROM tracked_targets AS tt INNER JOIN lifetime AS ltm ON (tt.date = ltm.date_first_seen) AND (tt.time >= ltm.time_first_seen) AND (tt.time <= ltm.time_last_seen); The preceding procedure returned close to 3,000 cases. These are possible near-crash events. To try to reduce the false positives, a filtering program was developed. The first step in the filtering program is to determine the predominate direction of travel for all the tracked vehicles. Essentially, this can be achieved by keeping track of the lane number as the target progresses in time. Second, since the primary interest is in events representing near-crash cases between a side-road vehicle and main-road vehicle, after the first sort, the side-road vehicles are examined. For each side- road vehicle, main-road vehicles are scanned at the same time step and run through a rejection process. In this process, if any test fails to return at least one possible conflicting main- road vehicle, the time step is advanced. The process is sum- marized in the following steps: Step 1: If no mainline vehicles are present for the given time step, then the time step is advanced. Step 2: A distance between the main-road vehicles and the side-road vehicle must lie below a specified threshold. Step 3: The speed of the side-road vehicle must be above a specified minimum threshold. Step 4: The speed of the main-road vehicle must be above a specified minimum threshold. The rejection at each step can be stored in a table that contains the date/time and target IDs. The level of elimina- tion depends on the specified speed/distance parameters of the trajectories. The result of the filtering program contained 297 cases of near crashes similar to a recorded crash case. A sample trajectory of a main-road tracked vehicle can be seen in Figure 5.2 for one of the resulting maneuvers. Note that as long as the lateral position relative to the center lane is within a certain margin, the target is snapped to the lane. Generally, what has been determined thus far is that the main-road tracked vehicle trajectories seem plausible. Unfortunately, this is not the case with the trajectories of the side-road vehicle. Although when the complete maneuver is plotted in space, the trajectories seem plausible (Figure 5.3), when animated in time, it becomes clear that peculiar tracking errors exist at critical locations dur- ing the maneuver. Figure 5.4 illustrates one such case, which has been typical of what has been observed thus far. The side-road tracked vehicle completed a northbound cross maneuver over US-74 (aligned east-west in Figure 5.3). It can be reasonably assumed that the vehicle never reverses course during the maneuver. Under such circumstances, the displace- ment curve from the start of the maneuver to the last tracked location should always monotonically increase. As seen in Figure 5.4, particularly at points A, B, and C, this is not the case. The side-road vehicle seems to make jumps backward, which is unrealistic. More important, these jumps do not seem to be Studied Intersection Figure 5.1. CICAS project at North Carolina site.
55 evidence is not enough to pinpoint the source of the difficulty and certainly not enough to allow the creation of automated correction methodologies, which, in any case, are outside the scope of this project. The aforementioned issues constrained the projectâs ability to process large numbers of possible near crashes. Regardless, the rest of this chapter includes the analysis, lessons learned, and modeling results for one near-crash event at the North Carolina CICAS site. CICaS North Carolina US-74e Near-Crash Case: 12:11 p.m., april 25, 2007 Lacking video for this selected event, the sequence of events can only be deduced from the animated trajectory informa- tion for the vehicles involved. Figure 5.5 shows a frame cap- tured from the visualization tool. isolated points on the overall trajectory, since the progression seems realistic following these course corrections. This last ele- ment renders difficult the correction of these trajectories even manually. After close inspection of many cases, numerous hypotheses have been formed on the nature of these discrepan- cies. Specifically, the errors seem to concentrate in areas where the side-road vehicle is stationary or a main-road vehicle is nearby. For the portable site deployments such as North Caro- lina, side-road vehicles are essentially tracked through the inter- section using horizontal LIDAR units placed at each approach of the side road and at a location in the median. The height of the LIDAR horizontal scan is set to approximate the height of the vehicle bumper. The LIDAR scanner sweep could have missed the bumper, occlusions from other vehicles, sensor alignment (as the tracked vehicle passes from one sensor data- collection field to another), or other algorithm processing fac- tors may also contribute to the error. Of course, such anecdotal Figure 5.2. Tracked target interface showing trajectories of a main-road vehicle.
56 Figure 5.3. Tracked target interface showing trajectories of a side-road vehicle. 0 50 100 150 200 250 300 0 5 10 15 20 25 30 Time in sec feet A B C Figure 5.4. Side-road vehicle displacements.
57 description of the event was deduced. Specifically, the side- road vehicle is performing a northbound crossing of the inter- section beginning from a standing position approximately 11 m (36 ft) from the conflict point. Considering this trajec- tory and the movement rate, the mainline vehicle, which initially is moving with an approximate speed of 26.8 m/s (60 mph), will eventually collide with the side-road vehicle, unless a change of attitude is implemented. The main-road driver, realizing this, performs a sequence of decelerations of up to more than -2 m/s2 (-6.6 ft/s2) to prevent the crash. The maneuver is performed successfully with room to spare. It is important to note that the described behavior does not con- sider any lateral movements of the vehicles, signifying types of evasive action other than deceleration, because if such movements did happen, they were smaller than the threshold used in the database for projecting the vehicle to the closest lane centerline. As discussed earlier, the animation of the side- road vehicle, which depended on the x, y coordinates, indi- cated that the reported speeds and accelerations do not agree. As seen in Figure 5.6, the side-road vehicle apparently never stops but appears to move with a constant speed of approxi- mately 11 mph (4.9 m/s) even while the acceleration increases from zero until well after the vehicle is clearly moving. There- fore, for the rest of the analysis, only the x, y coordinates of On the left side of Figure 5.5, there is a schematic of the intersection with the lines representing lane and turning cen- terlines, while on the right side the speed and acceleration graphs for the mainline and side-road vehicles are provided. A solid red line indicates the side-road vehicle, while magenta points indicate the mainline vehicle. On the graph side, the green vertical line indicates the values at the current frame displayed. In Figure 5.6, a zoomed view of the trajectories is presented. In Figure 5.6, the red circle indicates the mainline vehicle involved in the near crash. The vehicle moves toward the intersection in subsequent frames. The green (short) part of the side-vehicle trajectory is the movement already per- formed, and the blue (longer) part is the trajectory in future frames. The conflict point for these trajectories is also indi- cated. For the purposes of developing a model describing the vehicle trajectories, all coordinates were translated to a sys- tem that has its center at the conflict point and the x axis parallel to the mainline vehicle trajectory. This simplified the kinematic equations of the vehicles by using only the distance to the conflict point. The speed and acceleration values pro- vided in the CICAS database already represent the projected vectors on the trajectory of each vehicle. From the animation of the event, as well as the graphs of speed and acceleration, a Figure 5.5. US-74 near-crash event, 12:11 p.m., April 25, 2007.
58 cally, speed exhibits a single point drop of ~1.7 kph (1.0 mph), which, apart from being physically unrealistic, is not evident in the distance measurements. To explore the extent of this phenomenon, the project team proceeded to use the available data of speed and acceleration to estimate displacement per interval and compared this to the displacement produced by the x, y coordinates. This comparison is seen in Figure 5.8. For the purpose of the estimation, it was assumed that the starting values were known and the process progressed from there. The estimates of distance based on speed and accelera- tion, although not identical, show better agreement with each other than they do with the provided distance from the data- base. This can be seen as an indication that displacement uses the sensor range measurement, while speed and acceleration use range rate. Still, it is evident that significant filtering has been introduced. In Figure 5.9, a similar comparison is made on speed val- ues, one being provided and the other calculated based on the side-road vehicle are discussed, and speed and accelera- tion are inferred from them. The analysis of the data followed two routes. An empirical review of the information was performed to determine the concurrence of the provided information, followed by a more in-depth analysis producing Bayes estimates for each vehicleâs initial speed, the time points at which accelerations changed, and the accelerations in all stages. The reason for the empirical review was based on the quick realization of errors and dis- crepancies in the information provided, some of which was described earlier in this chapter and are illustrated in the fol- lowing sections specifically for this event. For the first part of the analysis, approximately 19 s of data were used for the mainline vehicle; 12 s of these were spent approaching the conflict point from a starting point approximately 276 m away. Figure 5.7 presents a subset of these for clarity. From this figure, certain problematic measurements, both in speed and distance, can immediately be observed. Specifi- Conflict Point Figure 5.6. Lane centerline trajectories and selected targets.
59 -950 -850 -750 -650 -550 -450 -350 -250 -150 -50 50 150 -14 -12 -10 -8 -6 -4 -2 2 0 Time in sec feet 0 10 20 30 40 50 60 70 80 90 100 feet/sec Distance in feet Speed in feet/sec Figure 5.7. Mainline vehicle distance and speed. -950 -850 -750 -650 -550 -450 -350 -250 -150 -50 -12 -10 -8 -6 -4 -2 0 2 Time in sec feet From speed From distance (provided) From acceleration Figure 5.8. Mainline vehicle distance comparison.
60 Exploratory modeling for both vehicles was conducted empirically in Excel to identify a plausible piecewise accelera- tion model. For the mainline vehicle, a three-stage model was fit, where a gentle deceleration lasting for about 3 s was fol- lowed by a stronger deceleration for 1 s and then by an even stronger one until the vehicle passed the conflict point. For the side-road vehicle, a one-stage model was fit, where a con- stant mild acceleration carried it over the first part of the intersection. To avoid some of the errors and data discrepan- cies discussed earlier, further analysis used a subset of the data starting at the point 8.2 s before the arrival of the main- line vehicle at the conflict point and ending at that point. Bayes estimates for each vehicleâs initial speed, the time points at which accelerations changed, and the accelerations in all stages were computed using WinBUGS. For this analy- sis, it was deemed necessary to remove some of the most obvious problems exhibited in the provided data. Specifi- cally, in the case of the mainline vehicle, the single point speed drop was corrected by adding a fixed value to all subse- quent data points. For the side-road vehicle, some abnormal back-and-forth data points at the beginning of the time period were removed. The WinBUGS results are displayed in Table 5.1. Figures 5.11 to 5.13 display the comparison between model estimates and provided measurements. One interesting observation resulting from the Bayes esti- mates of the driver behavior parameters involves the hypothet- ical mainline vehicle driver reaction time. Although the data are not sufficient to draw precise conclusions about the actions of the mainline driver, it can reasonably be assumed that the reac- tion to the side-road vehicleâs encroachment on the mainline acceleration. For the speed estimates calculated from accel- eration, the abnormal speed drop discussed earlier is not observed, but this can be either due to the introduction of an error in postprocessing or an artifact of the sampling meth- odology that will not allow it to propagate the acceleration measurements. That is, if accelerations were originally esti- mated as finite differences of speed, then a step change in speed will generally not be present in the acceleration esti- mates. Although in this case this single error can be easily corrected, further investigation of the sensor characteristics and data reduction methodology would be required to facili- tate better use of CICAS-collected mainline data with the trajectory-based methodology. Figure 5.10 presents the best information regarding the side-road vehicleâs distance-to-conflict point calculated from x, y coordinates. From the figure, it is evident that the mea- surements have both noise and errors and that these are not produced at the 10-Hz interval. The profile shows an unnatu- ral stepwise progress for the vehicle. In this figure, back-and- forth movements of the vehicle can be observed, as discussed earlier in this chapter. This case of a near crash was selected for analysis after a large number of events were reviewed to locate a relevant case with the fewest discrepancies, so that manual postprocessing with a minimal amount of abductive inference was possible. Specifically, in this case the problems can be found mainly in time periods that are not important for the analysis (before the vehicle comes to rest at the stop line and after it has sufficiently cleared the conflict point). Hence, further investigation used only information from time -6 s to time 1 s, approximately. 30 40 50 60 70 80 90 -12 -10 -8 -6 -4 -2 0 2 Time in sec Speed feet/sec Speed (provided) From acceleration Figure 5.9. Mainline vehicle speed comparison.
61 -90 -40 10 60 110 160 210 260 -12 -10 -8 -6 -4 -2 0 42 6 8 Time in sec feet Distance Figure 5.10. Side-road vehicle distance. Table 5.1. WinBUGS Estimates for CICAS Near-Crash Case Variable Mean Standard Deviation 2.5%tile 97.5%tile Side-Road Vehicle (initial speed zero) First acceleration (ft/s2) 4.74616 0.2188416 4.32632 5.18896 Movement start (s) -6.1422 0.102 -6.339 -5.95308 Main-Road Vehicle Initial Speed (ft/s) 82.0328 0.0498888 81.9344 82.1312 First acceleration (ft/s2) -2.4987 0.0270666 -2.551184 -2.44426 Second acceleration (ft/s2) -5.32016 0.0070487 -5.61536 -4.92984 Third acceleration (ft/s2) -7.04872 0.0201326 -7.0848 -7.00608 First change (s) -5.23692 0.03635 -5.31564 -5.17452 Second change (s) -4.323 0.06229 -4.45548 -4.22988
62 -50 -40 -30 -20 -10 0 10 20 30 40 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 Time in sec feet Distance provided Distance modeled corrected prior to fitting Figure 5.11. Measured and modeled side-road vehicle distances. 45 50 55 60 65 70 75 80 85 90 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 Time in sec Speed feet/sec Provided Modeled Figure 5.12. Measured and modeled main-road vehicle speeds.
63 of the video data was conducted manually; issues concerning automatic reduction of video data, although important, were treated as outside the scope of this project. The main objec- tive of the work described in Chapter 5 was to assess the abil- ity of an alternative technology, based on using Doppler shift methods, to support the analytic approach. One advantage of this alternative technology is that the problem of automati- cally extracting vehicle trajectories from video is avoided and that a database of potentially useful events currently exists. The project teamâs experience indicated that its structural model- ing and counterfactual screening methods can be applied to these data, but that to support this research, the existing data- collection and storage methods would require some technical modifications. must be somewhere between the first and second change time points, -5.24 to -4.323. If the Bayes estimate for the side vehicles starting time of encroachment is ignored and the data provided are followed literally, the encroachment may be said to start at around time point -5.012. This indicates a possible reaction time zone of -0.224 to 0.689 s. This is a tight reaction time for a driver noticing movement at a distance of more than 70 m (~230 ft). Alternatively, though, if the Bayes estimate for the side vehicle time of encroachment is accepted, then the reaction time zone is 0.9 to 1.8 s, which is much more reasonable. In conclusion, Chapter 4 described how the trajectory- based reconstruction method could be applied to site-based video data to estimate values of physical and behavioral vari- ables characterizing crash and near-crash events. Reduction -600 -500 -400 -300 -200 -100 0 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 Time in sec feet Distance modeled Distance provided Figure 5.13. Measured and modeled main-road vehicle distances.
