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.
2Overview Alcohol-impaired driving continues to be a significant public health concern. In 2011, 9,878 people were killed as a result of drunk driving in the United States, representing 31% of all traffic fatalities (1). Increased traffic crash risk results from drivers who experience pronounced and systematic physio- logical impairment while under the influence of alcohol (2, 3). Additionally, the intoxication level of passengers is linked to heightened crash risk through increased distraction of drivers (3, 4). Therefore, the presence of alcohol within a vehicle is a measure of significant interest when investigating crash causation. The second Strategic Highway Research Program (SHRP 2) Naturalistic Driving Study (NDS) has collected 50 million vehi- cle miles of naturalistic driving data on more than 3,000 drivers. The instrumentation of the vehicles used in this study included prototype alcohol-sensing technology that continuously detected the amount of ethanol in the vehicleâs interior. How- ever, this sensor outputs a raw value that is difficult for those unfamiliar with the technology to interpret. In addition, the presence of uncharacterized noise affects the sensorâs ability to detect the presence of alcohol. Several previous efforts explored the efficacy of the SHRP 2 alcohol sensor to detect the presence of imbibed alcohol. These included an experimental pilot study and Phase 1 (proof of concept) of the current project. The current effort, Phase 2, was built on previous attempts to develop an alcohol-detection algorithm that could be used to flag potential alcohol-involved cases in the SHRP 2 data set. Background Detailed descriptions of the SHRP 2 database, alcohol sensor, and previous research efforts follow. SHRP 2 Database The SHRP 2 database provided the primary data set used for early algorithm efforts and is also the primary database for this project. The SHRP 2 NDS is the largest naturalistic driv- ing study of its kind. As noted, it comprises 50 million miles of data, 5 million trip files, and more than 3,000 primary drivers. The data set contains video and kinematic information from each of the specially instrumented SHRP 2 vehicles. The abil- ity to detect the presence of alcohol using SHRP 2 sensors will help answer a multitude of research questions and will aid in evaluating the impact of alcohol on driver errors and crash likelihood. Hardware Configuration The alcohol sensor installed in the SHRP 2 vehicles was a model HS130D from Sencera Co., Ltd. (Figure 1.1). This sensor is a tin dioxide semiconductor gas sensor designed to quickly detect alcohol vapors at high relative humidity. The alcohol sensor was manufactured to detect the pres- ence of ethanol in the air. This could include alcohol vapors released into the cabin of a vehicle from the natural breathing of an individual who had consumed alcohol or a variety of unimbibed sources (e.g., hand sanitizer, perfume, mouthwash). The sensor returned a raw value in millivolts (mV). For the SHRP 2 project, the alcohol sensor was installed on the under- side of the head unit, which was mounted near the vehicleâs rearview mirror mount. This central location meant that it was equally able to detect alcohol vapors from the breathing of both the driver and the passenger. Pilot Investigation A pilot study of the alcohol sensor was conducted to examine the accuracy of the sensor under a variety of conditions. This C h a p t e r 1 Background
3 set. This was a limited exploratory effort that sought to develop an alcohol-detection algorithm from the SHRP 2 alcohol sen- sor and apply that algorithm to a selection of SHRP 2 trips. It also included limited sensor testing using a human partici- pant with the alcohol sensor in the same location as the final SHRP 2 sensor instrumentation. The human participant testing indicated that sensor read- ings in the SHRP 2 location decreased as a function of BrAC. Figure 1.2 shows the filtered alcohol sensor readings over time for occupants with differing BrAC levels. The vertical red lines represent the times when the windows were rolled up or down; the first red line marks windows being rolled down, the second red line shows windows being rolled up, and the third red line marks windows being rolled down again. This shows that alcohol sensor readings were sensitive to windows being rolled up or down and that sensors may, in fact, be able to detect the presence of alcohol because the change in sensor readings corresponds directly to the introduction of fresh air into the vehicle cabin. Despite the strong results, a successful algorithm was not developed in this initial phase that could reliably detect driver and passenger impairment in the SHRP 2 data set. This was partly due to the limited time, scope, and unanticipated chal- lenges that arose during this effort. For example, there was no true âgold standardâ data set of known alcohol-impaired trips to use as a model for alcohol sensor algorithm development. Also, a number of unanticipated substances were identified that seemed to strongly influence alcohol sensor readings (e.g., windshield wiper fluid and fast food). Given these out- comes, a second phase of research was initiated to continue development of the alcohol-detection method. Phase 2: Development and Evaluation of an Alcohol-Detection Algorithm The current effort, Phase 2, built on the results and over- came many of the challenges of the previous efforts. This research involved administering controlled doses of alcohol to researchers and having them sit in a vehicle equipped with an alcohol sensor. Researchers were administered alcohol to achieve target breath alcohol content (BrAC) readings of 0.05 grams per deciliter (g/dL) and 0.08 g/dL. These researchers then sat in either a front or back seat. Test conditions varied, based on whether the vehicle was moving, the air inside the vehicle was set to recirculate, or the windows were up or down. Results demonstrated that the prototype alcohol sensor configuration installed in the vehicle at that time could detect and distinguish varying levels of participant intoxication in several of these experimental conditions. In particular, the most accurate sensor readings occurred when the heating, ventilation, and air conditioning (HVAC) was set to recircu- late air and the windows were up. The sensor had difficulty detecting all but relatively high levels of alcohol when the HVAC system was on fresh air settings. Furthermore, the study also found differences in alcohol sensor readings for imbibed versus unimbibed alcohol. The primary difference was a steady, continually decreasing sensor reading for imbibed alcohol versus a sharper spike in sensor readings for unimbibed alcohol. These results suggested that a suitably designed algorithm could potentially differentiate between imbibed versus unimbibed alcohol. There were several significant changes between the alcohol sensor used in the pilot and the final SHRP 2 instrumentation. These include sensor position, the addition of a fan to circu- late air across the sensor in the SHRP 2 instrumentation, and significant software changes. Thus, while the results from the pilot research are useful, they cannot be generalized to alcohol sensor readings in the actual SHRP 2 data set. Phase 1: Alcohol-Detection Algorithm Proof of Concept The first phase of research examined the likelihood of devel- oping an alcohol-detection algorithm from the SHRP 2 data Figure 1.1. Alcohol sensor (left), sensor on circuit board (middle), and head unit as installed in SHRP 2 fleet (right).
4was primarily accomplished through the development of a gold standard data set through experimental in-vehicle test- ing and extensive data reduction of SHRP 2 trips for signs of impairment. The objectives of the effort were to develop 0 10 20 30 3600 3700 3800 3900 4000 4100 BrAC = 0.013 Fi lte re d al co ho l d at a 0 10 20 30 -4 -3 -2 -1 0 1 G ra di en t o f f ilt er ed a lc oh ol d at a 0 10 20 30 0 20 40 60 Sp ee d, m ph Time, minutes 0 10 20 30 3600 3700 3800 3900 4000 4100 BrAC = 0.031 0 10 20 30 -4 -3 -2 -1 0 1 0 10 20 30 0 20 40 60 Time, minutes 0 10 20 30 3600 3700 3800 3900 4000 4100 BrAC = 0.133 0 10 20 30 -4 -3 -2 -1 0 1 0 10 20 30 0 20 40 60 Time, minutes Figure 1.2. Alcohol sensor readings decrease as BrAC increases. an alcohol-detection algorithm to be applied to the SHRP 2 database and evaluate the accuracy of the algorithm at detect- ing and differentiating imbibed and unimbibed alcohol in the SHRP 2 database.
