Feasibility of Using In-Vehicle Video Data to Explore How to Modify Driver Behavior That Causes Nonrecurring Congestion (2011)

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.

C H A P T E R 4 Develop a Methodology for Analyzing DataA list of data elements has been identified from the quali- fied data sets in Chapter 3. From the various data sets, data elements that need to be considered could include video and vehicle kinematic data, in-vehicle and surrounding environ- mental data, vehicle- and infrastructure-based data, and raw and reduced data. Processing and analyzing the data requires addressing data storage issues, data storage configuration, reduction, and the computing necessary to analyze each data set. Because the candidate studies identified earlier are con- ducted by VTTI or UMTRI, the data computation capability of these two institutes is discussed briefly. Data Storage and Computation Requirements Naturalistic data collection studies that include video data usu- ally generate large files that require professional management. Consequently, research data at VTTI are stored on the Virtual Library System (VLS) Storage Area Network (SAN) that oper- ates within a dedicated private network. These data are isolated from VTTI’s operational network and all other networks, including the web, by high-end firewall hardware managed by VTTI’s Information Technology Group (ITG). All network connections within VTTI are high-speed Gigabit Ethernet. The data sets can be accessed using a dedicated, high-speed structured query language (SQL) server and by means of spe- cial application servers using Microsoft Sequel Server, MatLab, or SAS. Figure 4.1 shows the data center at VTTI. The data center has the following features: • Emergency power provided by an on-site diesel generator that supplies backup power for the data center, emergency lighting, and a telecommunications closet; • External wide area network (WAN) speeds equal to an OC-3 (45 Mbps, approximately 30 times that of a T-1 connection); • A dedicated climate control system with a backup contin- gency system;26• Remote monitoring alarm system for indication of fire, smoke, intrusion, power outage, climate control failure, hardware failure, water presence, and high temperature; • Elevated flooring system; • High physical security with steel-reinforced structural walls; • Limited personnel accessibility; and • Two 600-ft2 secure data reduction laboratories that house high-end Dell workstations. VTTI’s Data Services Center server room houses the fol- lowing equipment: • 250+ Dell-branded business-class desktops, laptops, and laboratory workstations; • 15+ high-availability, high-performance Dell PowerEdge servers; • More than 60 TB of redundant high-speed storage with short-term expandability exceeding 100 TB; • Redundant optical SAN switch/routers; • A large-capacity backup system that includes a tape library capable of handling 12 GB of data per min; and • Network connections that are all high-speed Gigabit Ether- net (VTTI was a pioneer in implementing this technology for every network portal). The computing requirements necessary for data manipu- lation and analysis are a result of the data size, data storage configuration, reduction needs, and analysis requirements. The VTTI team created codes that extract subsets of data for analysis purposes in multiple software environments, includ- ing MatLab, SAS, and SQL. An example of a flowchart for a MatLab function to extract a subset of data on a defined high- way section is provided in Figure 4.2. Besides the existing commercial software, VTTI developed proprietary data-viewing software, the Data Analysis and Reduction Tool (DART), to allow synchronized viewing of driver performance (parametric) data and video and audio

27Figure 4.1. Environmentally controlled and secured VTTI data center. Figure 4.2. Sample flowchart of a MatLab function.streams. This system allows researchers and data reductionists to work directly with large databases while providing ease of use and improved accuracy. As shown in Figure 4.3, reduction- ists can select specific variables and customize the interface of the illustration. While the video is playing, the software will draft charts for those variables along the time axis synchro- nized with the video. When multiple targets are sensed by radar units, the charts are color coded for better viewing.Similar to the data storage and calculation capability of VTTI, UMTRI has developed its data collection, storage, and computation capability over the years. UMTRI has devel- oped large, driver-vehicle databases since the mid-1990s. By the end of 2006, approximately 1 million vehicle miles of data had been collected. The data archive at UMTRI is maintained on an Internet-accessible system for on-demand access by

28Figure 4.3. DART developed by VTTI.registered users. Acquired data sets go through rigorous data quality assurance and validation procedures and can be refor- matted for use in standard analytical systems (e.g., ADAAS, SAS, and SPSS) and on multiple software platforms. Two servers hosted UMTRI data when it was first collected in 2004 through 2005. The first server is a Dell PowerEdge 4600 with a 12-GB system drive, a 546-GB RAID 10 drive containing data- base data files, and a 124-GB RAID 1 drive containing database log files. The server runs Windows 2003 Server SP2 and Microsoft SQL Server 2005. The video files are stored on a sep- arate Dell PowerEdge 4600 with a 1.2-TB logical drive. The servers are among several servers housing data from six FOTs and dozens of other projects that are located in a small, secure server room. UMTRI uses data structures that are intended to provide fast access so that custom tools allow users to access video in any project within 5 s. Direct access to the data and video is available for authorized users located within UMTRI, throughout the university, and within selected research part- ners, including access to all tools from remote locations. The descriptions provided summarize the data computation and storage abilities of VTTI and UMTRI. Data-processing capability does not have to achieve such levels to conduct video data–related research. Data storage and reduction will be satis-factory to make a judgment about nonrecurring congestion caused by driver behavior as long as the equipment can (1) secure the data so that only authorized personnel can access it; (2) be large enough to accommodate the video data, radar data, and associated vehicle and external data; (3) organize a relational database so that the data can be queried and retrieved efficiently; and (4) allow synchronized playback of the video data together with other vehicle and external data. Data Reduction and Crash and Near-Crash Detection As stated, data collection studies that include video data usually generate large files. The 100-Car Study raw data set, including both video and vehicle kinematic data, requires 7 TB of data storage. Reduced data sets require much less storage capacity. The quality of reduced data sets should be sufficient for the purposes of this study. With the exception of the NTNDS, which is still involved in an ongoing data reduction effort (see Table 3.7), the other three VTTI data sets and the two UMTRI FOT studies have been reduced and are discussed in detail in this chapter. The data reduction methodology, criteria to identify crashes and

29near crashes, and the associated weather and traffic condi- tions at the point that crashes and near crashes occurred are enumerated. Project 2: Automotive Collision Avoidance System Field Operational Test The original purpose of the study was to test the effectiveness of the FCW and ACC systems developed and integrated by GM and Delphi. The FCW delivers a warning when there is a for- ward crash possibility, and the ACC system maintains a safe headway. In addition to these systems, a DAS developed by UMTRI was installed on the test vehicles. The DAS deployed in this study organized data by trips. The system started to collect data each time the vehicle ignition was turned on and stopped when the engine was stopped. After each trip, files comprising the time history, transition, trip summary, triggered summary, raw CAN, video (each record contains a time and the bitmap image), and audio were recorded in the DAS. The video system had two cameras: a forward camera and a face camera. The forward scene was recorded continuously at 1 Hz, and the driver’s face was sampled at 5 Hz for 4 s every 5 min. Exposure video was saved simultaneously. One of the following triggers enabled the video system to capture a retro- spective clip, 8 s of video data (5 s before the event and 3 s after for the forward camera; 4 s before and 4 s after for the face camera), and saved the data to disk. These were activated by (1) a driver comment button; (2) Conventional Cruise Con- trol (CCC) and ACC engagement events; (3) FCW events; and (4) hard-braking events. The files recorded by the triggers were sent to UMTRI by way of cell modem for the purpose of preparing and facilitat- ing the participant debriefing. When the car was returned to UMTRI, software was used to process and view data, as shown in Figure 4.4 (1). Correspondingly, there were two phases of data reduction. The first phase was completed by the DAS as the vehicles were being driven by the participants. The onboard data-processing system generated CAN message lists and his- tograms for multiple variables, as well as an inventory file with all triggers. The second phase involved the processing of data by analysts after the vehicle was returned to UMTRI. The resulting data set of the ACAS FOT contained approximately 164 GB of data. Invalid trips were filtered out by discarding trips with any of the following attributes: zero distance; maximum speed less than 20 mph; system health histogram not okay; critical okay less than 90%; frozen threat index; radar malfunction; or non-FOT subject driving.exceed 2 m/s. Specifically, the weather was coded as 0 for “dry,” 1 for “wet,” and 2 for “snow covered.” As defined in the data reduction dictionary, the road was classified as wet if it was wet from snow but not snow covered. Also, any moisture on the road was classified as wet; there did not need to be standing water. A snow-covered classification would have included ice- covered if it was observed. If any portion of the road, including turn lanes, were covered in snow, the classification was snow covered. The traffic density was coded using the number of vis- ible traffic counts in the front camera, as shown in Figure 4.5 (1), in which “sparse” describes the top figure for which smoothed traffic count is less than 1.5 targets, “moderate” describes the middle figure for which smoothed traffic count is between 1.5 and 4.0 targets, and “dense” describes the bottom figure for which smoothed traffic count is greater than 4.0 tar- gets. The reduced data dictionary appears in Appendix A (1). Although the reduced data set included a description of secondary tasks (e.g., cell phone use, eating, drinking, smok- ing, and grooming), hand location, and eye location, it did not specify if the behavior was a contributing factor to the event. The team explored the possibility of conducting addi- tional data reduction to identify contributing factors but real- ized that the RDCWS FOT conducted by UMTRI (Project 5) has a similar but more sophisticated data structure compared with that for this project; that data set fits better with the cur- rent research goal. Also, the data transferring of this project has issues because of a waiver signed by the participants to allow for secondary usage of the data. To complete the addi- tional data reduction for the UMTRI studies in a timely man- ner, it was decided that the team would focus on the data collected from Project 5, and the results of data reduction performed by the team are discussed in Chapter 8.Project 5: Road Departure Crash Warning System Field Operational Test The research goal was to study the effectiveness and acceptance of the RDCWS, in which a Lane Departure Warning System (LDWS), a Curve Speed Warning System (CSWS), and a DAS were equipped onboard to provide warnings and collect data. The LDWS used a camera to observe visual features that delin- eate lane and road edges and a set of onboard radar units to modulate the warnings when potentially dangerous objects were sensed. The CSWS used GPS technology and a digital map to anticipate curve location and radius. The CSWS issued an alert when the speed was detected to be unsafe for the curve. The DAS, developed and managed by UMTRI, collected data from the RDCWS CAN bus, two radar buses, two video streams, an audio stream, and several other instruments. Data variables collected in this study are listed in Table 4.1 (2).While the video data were examined, environmental fac- tors (e.g., weather and traffic density), target characteristics, driver behavior, alert classification, and the driving scenario were identified and classified. A total of 634 alerts were man- ually examined wherever the target speed was estimated toThe two video cameras—one capturing the driver’s face and the other recording the forward view—recorded video

30Figure 4.4. Software interface used in Project 2.data at the frequencies shown in Table 4.2. Specifically, the DAS captured a 5-s video clip every 5 min from the face camera, regardless of the driving situations or driver activ- ity, creating a random sample of driver activity. The triggers (RDCWS alerts generated by the LDWS, CSWS, or a driver comment button pressed by the driver) sent a signal to the video system, and the cameras then recorded data at a higher frequency (as shown in Table 4.2) for 4 s before the event and 4 s after the alert.To reduce the data, analysts assigned values to 24 variables based on a review of the video and numerical data from the alerts. Twelve of the variables were associated with the cir- cumstances (e.g., road types and curves), and 12 variables addressing driver activities before, during, and after the alert (e.g., driver’s distractions, including glance direction) were coded for the events. Appendix B contains a summary of the data dictionary (2), and Figure 4.6 shows a screenshot of the video data collected in this project. Although the frequency

31Figure 4.5. Traffic density definitions in data reduction in Project 2.Table 4.1. Data Collected in Project 5 Data Sources Vehicle and driver identifications Vehicle position, heading, and motion: speed, yaw rate, accelerations, pitch and roll angle and rates, GPS (differential and nondifferential) Driver control inputs: steering wheel angle, throttle input, brake switch, turn signal, headlamp state, cruise control state and set speeds, LDWS and CSWS sensitivity settings RDCWS driver displays: LDWS and CSWS alerts and levels, availability icons RDCWS intermediate values: lane position, warning thresholds, road geometry estimates, threat locations, vehicle-centered object map Roadway environment: road type and attributes, urban and rural settings RDCWS and subsystem health and diagnostics information, as well as subsystem version numbers RDCWS radar data: forward radar data, side radar data Video: forward driving scene and driver face views Audio from driver comment button: dictated messages from driver Preset values CAN bus and FOT sensors CAN bus CAN bus CAN bus Onboard digital map via CAN bus, plus postprocessing, Highway Performance Monitoring System (HPMS) database CAN bus CAN bus LDWS camera, FOT sensors FOT sensors

32Table 4.2. Video Camera Configurations in Project 5 Nominal Triggered Rate Rate Pretrigger Posttrigger Video (Hz) (Hz) Window(s) Window(s) Forward video 2 10 4 4 Face video 0.5 10 4 4 Face video—exposure 5 n/a 0 5Figure 4.6. Video data collected in Project 5.set for the video cameras was relatively low, the data collected in this study are accurate, comprehensive, and efficiently reduced. Project 6: 100-Car Study The data reduction performed in this project can be summa- rized in two steps. First, events were identified using predefined trigger criteria values that resulted in a low miss rate and a high false alarm rate to avoid missing valid events. These criteria were derived after a sensitivity analysis. The specifics of the trigger values are described in Table 4.3. Second, the video data for all the events identified were reviewed by data reductionists. The reductionists focused on a 90-s epoch for each event (60 s before and 30 s after the event) to validate the event, determine severity, and code the event for a data reduction dictionary.conflicts were “any event that increases the level of risk asso- ciated with driving but does not result in a crash, near-crash, or incident. . . . Examples include: driver control error with- out proximal hazards being present; driver judgment error, such as unsafe tailgating or excessive speed; or cases in which drivers are visually distracted to an unsafe level.” Proximity events were “any circumstance resulting in extraordinarily close proximity of the subject vehicle to any other vehicle, pedestrian, cyclist, animal, or fixed object where, due to apparent unawareness on the part of the driver(s), pedestri- ans, cyclists or animals, there is no avoidance maneuver or response. Extraordinarily close proximity is defined as a clear case where the absence of an avoidance maneuver or response is inappropriate for the driving circumstances (including speed, sight distance, etc.).” Crash-relevant events were “any circumstance that requires a crash avoidance response on the part of the subject vehicle or any other vehicle, pedestrian, cyclist, or animal that is less severe than a rapid evasive maneu- ver, but greater in severity than a ‘normal maneuver’ to avoid a crash. A crash avoidance response can include braking, steering, accelerating, or any combination of control inputs.The severities of the valid events were determined based on various criteria, and all required variables were recorded and edited in MySQL databases. According to the severities, the valid events were classified into five categories: nonconflict, proximity events, crash-relevant, near crash, and crash. Non-

33Table 4.3. Event Triggers in Project 6 Trigger Type Description Lateral acceleration Longitudinal acceleration Event button Forward TTC Rear TTC Yaw rate Lateral motion equal to or greater than 0.7 g. Acceleration or deceleration equal to or greater than 0.6 g. Acceleration or deceleration equal to or greater than 0.5 g coupled with a forward TTC of 4 s or less. All longitudinal decelerations between 0.4 and 0.5 g coupled with a forward TTC value less than or equal to 4 s; corresponding forward range value at the minimum TTC is not greater than 100 ft. Activated by the driver pressing a button on the dashboard when an event occurred that he or she deemed critical. Acceleration or deceleration equal to or greater than 0.5 g coupled with a forward TTC of 4 s or less. All longitudinal decelerations between 0.4 and 0.5 g coupled with a forward TTC value less than or equal to 4 s; corresponding forward range value at the minimum TTC is not greater than 100 ft. Any rear TTC trigger value of 2 s or less that also has a corresponding rear range distance less than or equal to 50 ft and any rear TTC trigger value in which the absolute acceleration of the following vehicle is greater than 0.3 g. Any value greater than or equal to a plus and minus 4-degree change in heading (i.e., vehicle must return to the same general direction of travel) within a 3-s window.A ‘normal maneuver’ for the subject vehicle is defined as a control input that falls inside of the 99 percent confidence limit for control input as measured for the same subject.” Near crashes were defined as “any circumstance that requires a rapid, evasive maneuver by the subject vehicle, or any other vehicle, pedestrian, cyclist, or animal to avoid a crash. A rapid, evasive maneuver is defined as a steering, braking, accelerat- ing, or any combination of control inputs that approaches the limits of the vehicle capabilities. As a guide: subject vehicle braking greater than 0.5 g, or steering input that results in a lateral acceleration greater than 0.4 g to avoid a crash, consti-tutes a rapid maneuver.” Crashes were defined as events with “any contact with an object, either moving or fixed, at any speed, in which kinetic energy is measurably transferred or dissipated. Includes other vehicles, roadside barriers, objects on or off the roadway, pedestrians, cyclists or animals” (3–4). The variables defined in the reduced data can be catego- rized into one of the following homogeneous groups: general information, event variables, contributing factors, surround- ing factors, driver state variables, and driver information of the second vehicle involved. The variables are described in Table 4.4.Table 4.4. List of Variables in Reduced Data in Project 6 Classification List of Variables General information Event variables Contributing factors Environmental factors: driving environment Driving environment: infrastructure Driver state variable Driver/Vehicle 2 Vehicle number, epoch number, event severity, trigger type, driver subject number, onset of precipitating factor, and resolution of the event Event nature, incident type, pre-event maneuver, judgment of Vehicle 1 maneuver before event, precipitating fac- tor, evasive maneuver, and vehicle control after corrective action Driver behavior: Driver 1 actions and factors relating to the event, Driver 1 physical or mental impairment, Driver 1 distracted, willful behavior, driver proficiency, Driver 1 drowsiness rating, Driver 1 vision obscured by, and vehicle contributing factors Weather, light, windshield wiper activation, surface condition, and traffic density (level of service) Kind of locality, relation to junction, traffic flow, number of travel lanes, traffic control, and alignment Driver 1 hands on wheel, occupant safety belt usage, Driver 1 alcohol use, fault assignment, average PERCLOS, and Driver 1 eyeglance reconstruction Number of other vehicles/person(s), location of other vehicle(s)/persons, Vehicle/Person 2 type, Vehicle 2 maneu- ver, Driver/Vehicle 2 corrective action attempted, Driver/Vehicle 2 physical or mental impairment, and Driver 2 actions and factors relating to crash or incident

34As listed in Table 4.4, event nature was one of the variables coded in the reduced data set. Event nature usually was decided by reductionists based on predefined criteria together with subjective judgments. According to the definition, the valid events were classified into five categories according to their severity: nonconflict, proximity event, crash-relevant, near crash, and crash. These categories are described previously in this report. Among the safety-related events, crashes are relatively easy to identify. Near crashes and crash-relevant events are not as straightforward. Many safety applications would greatly ben- efit from a reliable, purely quantitative definition of a near crash and crash-relevant event. Because a near crash is more severe than a crash-relevant event and could have developed into a real crash, it is important to differentiate a near crash from a crash-relevant event. During data reduction, it was found that strictly applying a fixed number as a threshold value in acceleration or speed to identify near crashes inher- ently generates some “noise.” For example, a TTC of 0.1 s occurs regularly on interstates during congestion and is nor- mal, whereas the same TTC happening on a rural road at night is significantly different. Also, because of the varied driving habits of different drivers, it is hard to apply one uni- form number. Some drivers apply more frequent aggressive brakes simply because they are comfortable with rapid decel- erations. Therefore, both qualitative and quantitative criteria should be incorporated. As a guideline in this study, a subject vehicle braking greater than 0.5 g or a steering input that results in a lateral acceleration greater than 0.4 g to avoid a crash constitutes a candidate rapid maneuver, as defined in near crashes. Combined with the subjective judgment of experienced reductionists, this maneuver is decided by reduc- tionists as a near crash or crash-relevant event. A similar pol- icy for identifying near crashes was applied to most VTTI studies with appropriate adjustments made for threshold val- ues. Altogether, 69 crashes and 761 near crashes were identi- fied in the 100-Car data set. The crashes and near crashes were parsed into the following 18 conflict categories: • Conflict with a lead vehicle; • Conflict with a following vehicle; • Conflict with oncoming traffic; • Conflict with a vehicle in an adjacent lane; • Conflict with a merging vehicle; • Conflict with a vehicle turning across the subject vehicle path in the same direction; • Conflict with a vehicle turning across the subject vehicle path in the opposite direction; • Conflict with a vehicle turning into the subject vehicle path in the same direction; • Conflict with a vehicle turning into the subject vehicle path in the opposite direction;• Conflict with a vehicle moving across the subject vehicle path through the intersection; • Conflict with a parked vehicle; • Conflict with a pedestrian; • Conflict with a pedal cyclist; • Conflict with an animal; • Conflict with an obstacle or object in the roadway; • Single-vehicle conflict; • Other (specify); and • Unknown conflict. Table 4.5 shows the level of service (LOS) at the time the crash or near crash happened. Preferably, the LOS both before and after the event can be identified so that the travel time reliability caused by incidents can be analyzed and mod- eled. (Details of modeling travel time reliability are discussed in Chapter 6.) Once a car was involved in an accident and the engine was turned off, the onboard data collection system would be turned off and would cease collecting data. There- fore, the LOS after events usually was not available. Table 4.5 lists only the LOS before the events occurred. The weather conditions associated with safety-related events varied. Clear weather was associated with the same percentage of 100-Car Study crashes (78%) and near crashes (78%). The second most associated weather factor was rain. Drivers had a slightly higher percentage of crashes (12%) than near crashes (8%) that occurred during rainy days. Cloudy weather was associated with more near crashes (13%) than crashes (9%). Only one 100-Car Study crash occurred with snow as an associated factor.Project 7: Drowsy Driver Warning System Field Operational Test Data reduction started with the identification of potential events using the DART. A 90-s epoch was created for each event, which included 1 min before the trigger and 30 s after. The automatic scanning resulted in an event database. The triggers and values used in identifying critical events are listed in Table 4.6.Data reductionists reviewed the video data for the identi- fied events to validate them. Invalid events for which sensor readings were spurious because of a transient spike or some other false-positive classification were filtered out. Valid events were classified as conflicts or nonconflicts. Conflicts were further classified into four categories based on the sever- ity: crash, crash: tire strike, near crash, and crash-relevant. Non- conflicts were events with valid threshold values but did not create safety-significant traffic events. Verified valid conflicts were categorized based on the following descriptions. Crashes are classified as any contact with an object, either moving or fixed, at any speed in which kinetic energy is measurably

35Table 4.5. LOS for Crashes and Near Crashes by Conflict Type Crashes Near Crashes Conflict type LOS A LOS B LOS C+ LOS A LOS B LOS C+ Single 24 20 3 1 48 42 3 3 Lead 15 5 5 5 380 78 241 61 Following 12 4 7 1 70 16 39 15 Obstacle 9 6 3 0 6 1 5 0 Animal 2 2 0 0 10 10 0 0 Turning across opposite direction 2 1 1 0 27 12 15 0 Adjacent vehicle 1 0 1 0 115 32 69 14 Parking 4 3 1 0 5 3 2 0 Across path through intersection NA 27 11 15 1 Oncoming 27 14 12 1 Other 2 1 1 0 Pedestrian 6 4 1 1 Turning across in same direction 3 2 1 0 Turning in same direction 28 16 12 0 Merging 6 2 4 0 Unknown 1 1 0 0Table 4.6. Trigger Values to Identify Critical Incidents in Project 7 Trigger Type Description Longitudinal acceleration (LA) Time to collision (TTC) Swerve (S) Critical incident (CI) button Analyst identified (AI) (1) Acceleration or deceleration greater than or equal to⎟ 0.35 g⎟ . Speed greater than or equal to 15 mph. (2) Acceleration or deceleration greater than or equal to⎟ 0.5 g⎟ . Speed less than or equal to 15 mph. (3) A forward TTC value of less than or equal to 1.8 s, coupled with a range of less than or equal to 150 ft, a target speed of greater than or equal to 5 mph, a yaw rate of less than or equal to⎟ 4°/s⎟ , and an azimuth of less than or equal to⎟ 0.8°⎟ . (4) A forward TTC value of less than or equal to 1.8 s, coupled with an acceleration or deceleration greater than or equal to⎟ 0.35 g⎟ , a forward range of less than or equal to 150 ft, a yaw rate of less than or equal to⎟ 4°/s⎟ , and an azimuth of less than or equal to⎟ 0.8°⎟ . (5) Swerve value of greater than or equal to 3. Speed greater than or equal to 15 mph. (6) Activated by the driver pressing a button, located by the driver’s visor, when an incident occurred that he or she deemed critical. (7) Event identified by a data reductionist viewing video footage; no other trigger listed above identified the event (e.g., LA and TTC).transferred or dissipated. Near crashes (evasive maneuver) are classified as any circumstance that requires a rapid, eva- sive maneuver, including steering, braking, accelerating, or any combination of control inputs that approaches the limits of the vehicle capabilities, by the subject vehicle or any other vehicle, pedestrian, cyclist, or animal to avoid a crash. Near crashes (no evasive maneuver) are classified as any circum- stance that results in extraordinary proximity of the subjectvehicle to any other object. Extraordinary proximity is defined as a clear case in which the absence of an avoidance maneuver or response is inappropriate for the driving circumstances (e.g., speed and sight distance). Crash-relevant conflicts (eva- sive maneuvers) are assessed similar to near crashes (evasive maneuvers). Longitudinal decelerations of −0.35 g or greater are reviewed to assess whether they qualify as crash-relevant con- flicts (or near crashes); those with decelerations of −0.50 g or

36greater are always coded as crash-relevant conflicts or near crashes. Crash-relevant conflicts (no evasive maneuver) are classified similar to near crashes (no evasive maneuver). Here a TTC of 1 s is used to identify near crashes. Cases with a TTC between 1 and 2 s relied on subjective judgment (5). Relevant information was coded in a data dictionary that includes more than 50 variables (Appendix C). Figure 4.7 provides the flowchart showing the process of data reduction. After data reduction, 915 safety-relevant events were identi- fied; of these, there were 14 crashes, 14 tire strikes, 98 near crashes, and 789 crash-relevant conflicts. A random sample of 1,072 baseline epochs was also selected to represent normal driving. Each baseline epoch is 60 s long. The data dictionary was also applied to those baseline epochs. The purpose of select- ing baseline events was to provide a comparison with the events in that the baseline events represent normal driving. Figure 4.8 shows a screenshot of the video data in this study (5).Events Database Valid Events? No Yes Conflicts? Yes Classified and reduced in data dictionary: • Crash • Crash: Tire strike • Near crash • Crash-relevant conflict No End Start Data Scanning Figure 4.7. Flowchart for data reduction in Project 7.Figure 4.8. Screenshot from Project 7.venience provided to the motorist, passenger, or pedestrian is excellent; 2. LOS B/Flow with some restrictions: In the range of stable traffic flow, but the presence of other users in the traffic stream begins to be noticeable. Freedom to select desired speeds is relatively unaffected, but there is a slight decline in the freedom to maneuver within the traffic stream from LOS A because the presence of others in the traffic stream begins to affect individual behavior; 3. LOS C/Stable flow: Maneuverability and speed are more restricted. In the range of stable traffic flow but marks the beginning of the range of flow in which the operation of individual users becomes significantly affected by the inter- actions with others in the traffic stream. The selection of speed is now affected by the presence of others, and maneu- vering within the traffic stream requires substantial vigi- lance on the part of the user. The general level of comfort and convenience declines noticeably at this level; 4. LOS D/Unstable flow: Temporary restrictions substantially slow driver. Represents high-density and unstable traffic flow. Speed and freedom to maneuver are severely restricted, and the driver or pedestrian experiences a generally poor level of comfort and convenience. Small increases in traffic flow will generally cause operational problems at this level; 5. LOS E: Vehicles are unable to pass and there are temporary stoppages. Represents operating conditions at or near the capacity level. All speeds are reduced to a low but relatively uniform value. Freedom to maneuver within the traffic stream is extremely difficult, and it is generally accom- plished by forcing a vehicle or pedestrian to yield to accom- modate such maneuvers. Comfort and convenience levels are extremely poor, and driver or pedestrian frustration is generally high. Operations at this level are usually unstable because small increases in flow or minor perturbations within the traffic stream will cause breakdowns;In characterizing traffic density at the time the event hap- pened, a detailed description was provided to the data reduc- tionists to assist in assigning the LOS. According to the report from Project 7 (5), six levels of traffic density are defined plus a status of unknown or unable to determine: 1. LOS A/Free flow: Individual users are virtually unaffected by the presence of others in the traffic. Freedom to select desired speeds and to maneuver within the traffic stream is extremely high. The general level of comfort and con-

376. LOS F: Forced traffic flow condition with low speeds and traffic volumes that are below capacity; queues form in par- ticular locations. This condition exists whenever the amount of traffic approaching a point exceeds the amount that can traverse the point. Queues form behind such locations. Operations within the queue are characterized by stop-and- go waves, and they are extremely unstable. Vehicles may progress at reasonable speeds for several hundred feet or more and then be required to stop in a cyclic manner. LOS F is used to describe the operating conditions within the queue, as well as the point of the breakdown. In many cases, oper- ating conditions of vehicles or pedestrians discharged from the queue may be quite good. It is the point at which arrival flow exceeds discharge flow that causes the queue to form, and LOS F is an appropriate designation for such points. Table 4.7 shows the numbers and percentages of crashes, crashes: tire strikes, and near crashes associated with each LOS.Table 4.7. Level of Service for Crashes and Near Crashes No. of % of No. of % of No. of % of Crashes: Crashes: Near Near Traffic Density Crashes Crashes Tire Strikes Tire Strikes Crashes Crashes LOS A 13 92.9% 9 64.3% 61 62.2% LOS B 1 7.1% 1 7.1% 21 21.4% LOS C 0 0.0% 4 28.6% 11 11.2% LOS D 0 0.0% 0 0.0% 1 1.0% LOS E 0 0.0% 0 0.0% 2 2.0% LOS F 0 0.0% 0 0.0% 2 2.0% Unknown 0 0.0% 0 0.0% 0 0.0% Total 14 100.0% 14 100.0% 98 100.0%Weather conditions were coded as “No adverse conditions,” “Rain,” “Sleet,” “Snow,” “Fog,” “Rain and fog,” “Sleet and fog,” “Other,” and “Unknown.” Table 4.8 shows the numbers and percentages of crashes, crashes: tire strikes, and near crashes associated with each weather condition.Table 4.8. Weather Condition When Events Happened No. of % of No. of % of No. of % of Crashes: Crashes: Near Near Weather Crashes Crashes Tire Strikes Tire Strikes Crashes Crashes No adverse conditions 11 78.6% 14 100.0% 91 92.9% Rain 2 14.3% 0 0.0% 7 7.1% Sleet 0 0.0% 0 0.0% 0 0.0% Snow 0 0.0% 0 0.0% 0 0.0% Fog 1 7.1% 0 0.0% 0 0.0% Rain and fog 0 0.0% 0 0.0% 0 0.0% Sleet and fog 0 0.0% 0 0.0% 0 0.0% Other 0 0.0% 0 0.0% 0 0.0%Project 8: Naturalistic Truck Driving Study Data reduction for the NTDS involved two main steps. Step 1 was to identify events of interest. The DART was used to find events of interest by scanning the data set for notable actions, including hard braking, quick steering maneuvers, short TTCs, and lane deviations. Table 4.9 displays the vari- ous trigger threshold values (6). VTTI researchers developed the values based on data reduction experience obtained from the 100-Car Study. A 75-s epoch was created for each trigger comprising 1 min before the trigger and 15 s after the trigger. The result of the automatic scan was an event data set that

38Table 4.9. Trigger Values Used to Identify Critical Incidents in Project 8 Trigger Type Definition Description Longitudinal acceleration (LA) Time to collision (TTC) Swerve (S) Lane deviations (LD) Critical incident (CI) button Analyst identified (AI) Hard braking or sudden acceleration Amount of time (in seconds) it would take for two vehicles to collide if one vehicle did not perform an evasive maneuver. Sudden jerk of the steering wheel to return the truck to its original position in the lane. Any time the truck aborts the lane line and returns to the same lane without making a lane change. Self-report of an incident by the driver. Event identified by the analyst but not by a trigger. Acceleration or deceleration greater than or equal to⎟ 0.20 g⎟ . Speed greater than or equal to 1 mph (1.6 km/h). A forward TTC value of less than or equal to 2 s, coupled with a range of less than or equal to 250 ft, a target speed of greater than or equal to 5 mph (8 km/h), a yaw rate of less than or equal to⎟ 6°/s⎟ , and an azimuth of less than or equal to⎟ 0.12°⎟ . Swerve value of greater than or equal to 2 rad/s2. Speed greater than or equal to 5 mph (8.05 km/h). Lane tracker status equals abort. Distance from center of lane to outside of lane line less than 44 in. Activated by the driver pressing a button by the driver’s visor when an incident occurred that he or she deemed critical. Event that was identified by a data analyst viewing video footage; no other trigger listed above identified the event (e.g., LA and TTC).included both valid and invalid events waiting to be further identified in step 2. Step 2 involved a manual inspection of these potential events of interest by data reductionists to filter out invalid events. Figure 4.9 shows a screenshot of the video data. Valid events were further classified into one of six safety- critical events: crash, crash: tire strike, near crash, crash- relevant conflict, unintentional lane deviation, and illegal maneuver. Table 4.10 summarizes the definitions of these event types (6).Figure 4.9. Screenshot of video data in Project 8.ated video data and answering questions in a pull-down menu in the DART. Invalid events were eliminated when sen- sor readings were spurious because of a transient spike or some other anomaly (i.e., false positive). Appendices C and D provide the data dictionary for event and environmental cri- teria, respectively. Most of the events happened during smooth traffic condi- tions and nonadverse weather conditions. Tables 4.11 and 4.12 show the details of the LOS and weather categories, respec- tively. The traffic and weather conditions at the instant the crashes and near crashes occurred should not be used alone to decide if these factors have an impact on the likelihood of the safety-related events. The percentage of heavy traffic and adverse weather conditions occurring in baseline epochs should be compared with the event epochs. For example, heavy traffic occurs in 10% of the baseline epochs but is present in 20% of the event epochs. Although 20% is not a high percent- age, it is a factor worthy to note.Project 11: Naturalistic Teen Driving Study Data reduction procedures in this study included three tasks: initial data reduction, straight road segment data reduction, and event data reduction. The initial data reduc- tion involved recording general information about the driver and passengers of each trip. A trip was defined as a con- tinuous data collection from the start of the engine of a partic- ipant’s vehicle to its turn-off. The recorded variables included participant ID, number of passengers, age of passengers,The details of each valid event were coded by reductionists using an established coding directory by watching the associ-

39Table 4.10. Event Types in Project 8 Event Type Description Crash Crash: tire strike Near crash Crash-relevant conflict Unintentional lane deviation Illegal maneuver Any contact with an object, either moving or fixed, at any speed. Any contact with an object, either moving or fixed, at any speed in which kinetic energy is measurably trans- ferred or dissipated when the contact occurs only on the truck’s tire. No damage occurs during these events (e.g., a truck is making a right turn at an intersection and runs over the sidewalk or curb with a tire). Any circumstance that requires a rapid, evasive maneuver (e.g., hard braking, steering) by the subject vehicle or any other vehicle, pedestrian, cyclist, or animal to avoid a crash. Any circumstance that requires a crash-avoidance response on the part of the subject vehicle, any other vehi- cle, pedestrian, cyclist, or animal that was less severe than a rapid evasive maneuver (as defined above) but more severe than a normal maneuver. A crash-avoidance response can include braking, steering, accelerating, or any combination of control inputs. Any circumstance in which the subject vehicle crosses over a solid lane line (e.g., onto the shoulder) where there is not a hazard (guardrail, ditch, or vehicle) present. Any circumstance in which either the subject vehicle or the other vehicle performs an illegal maneuver, such as passing another vehicle across the double yellow line or on a shoulder. In many cases, neither driver performs an evasive action.Table 4.11. LOS for Crashes and Near Crashes in Project 8 No. of % of No. of % of No. of % of Crashes: Crashes: Near Near Traffic LOS Crashes Crashes Tire Strikes Tire Strikes Crashes Crashes LOS A 3 60.0% 4 50% 23 37.7% LOS B 2 40.0% 3 37.5% 22 36.1% LOS C 0 0.0% 0 0.0% 13 21.3% LOS D 0 0.0% 0 0.0% 1 1.6% LOS E 0 0.0% 1 12.5% 2 3.3% LOS F 0 0.0% 0 0.0% 0 0.0% Unknown 0 0.0% 0 0.0% 0 0.0% Total 5 100.0% 8 100.0% 61 100.0%Table 4.12. Weather Condition for Crashes and Near Crashes in Project 8 No. of % of No. of % of No. of % of Crashes: Crashes: Near Near Weather Crashes Crashes Tire Strikes Tire Strikes Crashes Crashes No adverse conditions 5 100.0% 8 100.0% 56 91.8% Rain 0 0.0% 0 0.0% 5 8.2% Sleet 0 0.0% 0 0.0% 0 0.0% Snow 0 0.0% 0 0.0% 0 0.0% Fog 0 0.0% 0 0.0% 0 0.0% Rain and fog 0 0.0% 0 0.0% 0 0.0% Sleet and fog 0 0.0% 0 0.0% 0 0.0% Other 0 0.0% 0 0.0% 0 0.0% Total 5 100.0% 8 100.0% 61 100.0%

40and time of day. For the straight road segment reduction, data reductionists recorded a set of variables for 22 previ- ously selected straight road segments that were chosen based on the frequency of travel. The recorded variables included driver behavior and performance in an attempt to analyze driver engagement in secondary tasks and the driver’s eye- scanning patterns. For the event reduction, potential events were initially identified by the kinematic trigger values, and trained reductionists reviewed the video data to confirm their validity. The details of the trigger values are described in Table 4.13. Valid events were further classified by data reductionists into three categories—crash, near crash, and judgment error— based on their severity. A crash is defined as “any contact with an object, either moving or fixed, at any speed in which kinetic energy is measurably transferred or dissipated. Includes other vehicles, roadside barriers, objects on or off the road- way, pedestrians, cyclists, or animals.” A near crash is defined as “any circumstance that requires a rapid evasive maneu- ver by the subject vehicle or any other vehicle, pedestrian, cyclist, or animal to avoid a crash. A rapid evasive maneuver is defined as steering, braking, accelerating, or any combina- tion of control inputs that approaches the limits of the vehi- cle capabilities.” Judgment error is defined as “any circumstance where the teen driver purposefully or inadvertently creates a safety- relevant situation due to either general inexperience or per- formance error. Examples include those events where the teen drivers engage in ‘horseplay’ or overreact to surround- ing traffic” (7). For crash and near crash events, a series of variables were coded during the data reduction process. The variables are broadly classified into several homogeneous groups as vehicle, event, environmental, driver’s state, and second vehicle vari- ables, as summarized in Table 4.14.Table 4.13. Trigger Types in Project 11 Trigger Type Description Longitudinal deceleration (LD) Lateral acceleration Forward time to collision (FTTC) Yaw rate Longitudinal acceleration (LA) Critical incident (CI) button Speeding trigger Less than or equal to −0.65 g Greater than or equal to 0.75 g or less than or equal to −0.75 g • FTTC less than or equal to 4 s with a deceleration less than or equal to −0.6 g. • FTTC less than or equal to 4 s paired with a deceleration less than or equal to −0.5 g and forward range of less than or equal to 100 ft. Vehicle swerves ±4 degrees/s to ±4 degrees/s within a window of 3.0 s Greater than or equal to 0.5 g, returning to 0.1 g within 0.2 s Boolean response In excess of 70 mph but not traveling on an interstateTable 4.14. List of Variables in Reduced Data in Project 11 Classification Examples Vehicle variables Event variables Environmental variables Driver’s state Driver of vehicle 2 Vehicle ID, vehicle type, owned or shared, and VMT Nature of event or crash type, pre-event maneu- ver, precipitating factors, corrective action or evasive maneuver, contributing factors, types of inattention, driver impairment Weather, ambient lighting, road type, traffic density, relation to junction, surface condi- tion, traffic flow Hands on wheel, seat belt usage, fault assign- ment, eyeglance Vehicle 2 body style, maneuver, corrective action attemptedFigure 4.10 illustrates some sample screenshots for the NTNDS. The left screenshot is a snapshot from the continu- ous video data captured by four continuous camera views monitoring the driver’s face and the driver side of the vehicle, the forward view, the rear view, and an over-the-shoulder view for the driver’s hands and surrounding areas. In the right screenshot, the bottom photos show sample periodic still shots by two other cameras in the interior vehicle cabin, as well as the lap area of the rear passenger seat.Data reduction for this study is ongoing, and not much can be presented in this report. Because this study is the most recent naturalistic study accomplished by VTTI, all the equipment involved is the most updated and highly accurate. The data reduction protocol is refined so that the threshold values selected are more reasonable, and data reductionists have acquired more experience from previous studies. The resulting data set will be beneficial in the next stage of research for this study.

41Figure 4.10. Screenshots of video data in Project 11.References 1. University of Michigan Transportation Research Institute. Automotive Collision Avoidance System Field Operational Test Report: Methodology and Results. Report DOT HS 809 900. NHTSA, 2005. 2. University of Michigan Transportation Research Institute. Road Departure Crash Warning System Field Operational Test: Methodology and Results. NHTSA, 2006. 3. Dingus, T. A., S. G. Klauer, V. L. Neale, A. Petersen, S. E. Lee, J. Sudweeks, M. A. Perez, J. Hankey, D. Ramsey, S. Gupta, C. Bucher, Z. R. Doerzaph, J. Jermeland, and R. R. Knipling. The 100-Car Nat- uralistic Driving Study, Phase II: Results of the 100-Car Field Experi- ment. Report DOT HS 810 593. NHTSA, 2006. 4. Neale, V. L., S. G. Klauer, R. R. Knipling, T. A. Dingus, G. T. Holbrook, and A. Petersen. The 100-Car Naturalistic Driving Study, Phase 1: Experimental Design. Report DOT HS 808 536. NHTSA, 2002.5. Hanowski, R. J., M. Blanco, A. Nakata, J. S. Hickman, W. A. Schaudt, M. C. Fumero, R. L. Olson, J. Jermeland, M. Greening, G. T. Holbrook, R. R. Knipling, and P. Madison. The Drowsy Driver Warning System Field Operational Test, Data Collection: Final Report. Report DOT HS 811 035. NHTSA and Virginia Tech Transportation Institute, Blacksburg, Va., 2005. 6. Blanco, M., J. S. Hickman, R. L. Olson, J. L. Bocanegra, R. J. Hanowski, A. Nakata, M. Greening, P. Madison, G. T. Holbrook, and D. Bowman. Investigating Critical Incidents, Driver Restart Period, Sleep Quantity, and Crash Countermeasures in Commercial Vehicle Operations Using Naturalistic Data Collection. FMCSA, 2008. 7. Lerner, N., J. Jenness, J. Singer, S. G. Klauer, S. Lee, M. Donath, M. Manser, and M. Ward. An Exploration of Vehicle-Based Monitor- ing of Novice Teen Drivers: Draft Report. Virginia Tech Transporta- tion Institute, Blacksburg, Va., 2008.