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.

Nonrecurring congestion is traffic congestion due to nonrecurring causes, such as crashes, dis- abled vehicles, work zones, adverse weather events, and planned special events. According to data from the Federal Highway Administration (FHWA), approximately half of all congestion is caused by temporary disruptions that remove part of the roadway from use, or “nonrecurring” congestion. These nonrecurring events dramatically reduce the available capacity and reliability of the entire transportation system. The objective of this project is to determine the feasibility of using in-vehicle video data to make inferences about driver behavior that would allow investi- gation of the relationship between observable driver behavior and nonrecurring congestion to improve travel time reliability. The data processing flow proposed in this report can be summa- rized as (1) collect data, (2) identify driver behavior, (3) identify correctable driver behavior, and (4) model travel time reliability, as shown in Figure ES.1. Executive SummaryIntroduction Key domestic and international studies in which in-vehicle video cameras were used to collect data were investigated in this study. The research team reviewed video, kinematic, and external data col- lected in each candidate data set. To quantitatively assess the qualification of candidate data sets, dimensions of feasibility were defined to evaluate the data sets with respect to legal restrictions, comprehensiveness, video data quality, in-vehicle data quality, linkages to external data, and format and structure. A list of qualified data sets was identified for further data reduction and analysis. The original research goals, data reduction process, and data formats of these studies were examined by the research team. The video data were manually reviewed, and additional data reduction was conducted to identify contributing factors to crashes and near crashes using video data and supplementary data. If the events were caused by inappropriate driver behavior or driver inattention, then countermeasures were suggested to prevent these safety-related events. In modeling travel time reliability, the team reviewed models used by other researchers. A multimode distribution of travel time was then proposed to model travel time variations. A sta- tistical method to model travel time reliability was developed that considers two factors: the probability of encountering congestion and the probability that certain estimated travel times will be experienced when congestion occurs. Potential problems and risks associated with existing data, including kinematic data, video data, reduced data, and external data, were identified during the data reduction and analysis phase. To facilitate future data collection efforts, including those related to in-vehicle data, other external data were proposed to improve the efficiency of data collection, reduction, and analysis.1

2Figure ES.1. Data processing flow. FCW = forward crash warning.Findings According to data reduction results, most crashes or near crashes are caused by driver inatten- tion and errors. These events might be prevented if appropriate instrumentation were installed to warn drivers in a timely manner. The following factors imply that the proposed systems can resolve driver inattention and errors and thus prevent a collision: 1. In the Road Departure Crash Warning System (RDCWS) Field Operational Test (FOT), the largest contributing factor to freeway crashes and near crashes was decision errors, including driving too fast or too slowly, following too closely, and misjudging a gap; more than 85% of the events were caused by this driver-related factor. For events that occurred on arterials, the same pattern followed. The next largest category for both road types was recognition errors, including inattention, inadequate surveillance, and other types of distraction; more than 5% of events were ascribed to this category. 2. In the 100-Car Study, the largest contributing factor category for crashes was driver recogni- tion errors, totaling 32% of the events. The second largest category was decision errors, count- ing 28% of the total. The largest and second largest contributing factor categories for near crashes were decision errors and recognition errors at 29% and 26%, respectively. 3. In the Drowsy Driver Warning System (DDWS) FOT, a different pattern was exhibited. The most frequent critical reason for crashes was an object in the roadway, which constituted 57% of the events. The next largest groups were the driver-related factors of recognition errors and performance errors; each had more than 14% of the cases related to driver factors. In tire strike cases, the majority were attributed to environment-related factors; more than 64% of the events were ascribed to this category. For near crashes, recognition errors and decision errors constituted 31% and 18%, respectively. 4. In the Naturalistic Truck Driving Study (NTDS), the most frequent critical factor for crashes was an object in the roadway, followed by the driver-related factors of recognition errors, deci- sion errors, and performance errors; each constituted 20% of the total cases. Not surprisingly, almost all (75%) the tire strikes involved some type of improper turn. The second and third largest categories of contributing factor for crashes were driver performance errors and deci- sion errors, respectively. For near crashes, the most frequent factor was driver-related recog-

3nition errors; more than 40% of near crashes resulted from inattention or distraction. Among these near crashes, almost one-quarter involved the subject driver’s not seeing the other vehicle during a lane change or merge. Countermeasures were proposed to correct the driver behaviors that caused these events. Almost all the crashes in the RDCWS FOT study have the potential to be prevented if one or multiple countermeasures are applied; 91% of the near crashes in that study are correctable. In the 100-Car Study, almost 40% of the crashes can or are likely to be prevented, and more than 80% of the near crashes can or are likely to be prevented given reasonable countermeasures. In the two truck studies, all the crashes, tire strikes, and near crashes are preventable using appro- priate countermeasures. To model nonrecurring congestion related to modifying driver behavior, it is ideal to find a sub- stantial number of crashes that result in changes in traffic conditions. The congestion, therefore, can be monitored and modeled. According to the data reduction results, not enough realizations of such crashes occurred. Other supplemental data sets were used to construct the statistical model. The travel time data in the 100-Car Study were used only to validate the multimode distribution of travel time that is being proposed; the model is valid regardless of the source of the travel time data. The travel time reliability model results provide a better fit to field data compared with tra- ditional unimodal travel time model results. The reliability model is also more flexible and pro- vides superior fitting to travel time data compared with single-mode models. It provides a direct connection between the model parameters and the underlying traffic conditions and can be directly linked to the probability of incidents. Thus, it can capture the impact of nonrecurring congestion on travel time reliability. Conclusions The team explored the identified data sets to discuss the various issues associated with video and other supplementary data collection and data reduction, to propose models for travel time relia- bility, and to identify potential problems in data. From the analysis of the naturalistic data sources, this study demonstrates the following: 1. It is feasible to identify driver behavior before near crashes and crashes from video data col- lected in a naturalistic driving study and to thus infer the causes of those events. 2. Recommendations can be made to change driver behavior and, therefore, prevent crashes and near crashes or reduce the frequency of such events. 3. Naturalistic data are useful to identify impacts of crashes on traffic conditions. Given the small sample of crashes and the fact that the data acquisition system (DAS) does not gather data when the engine is off, it is not possible to study the impact of incidents on travel time reliability. When effectively integrated with external data sources, which is extremely feasi- ble given an accurate time and location stamp in the data set, naturalistic data can be highly efficient in recognizing the relationship between modifying driver behavior and nonrecur- ring congestion. 4. Increased coordination with weather and traffic volume data is required to determine when nonrecurring congestion exists, as well as to determine what driver actions are a result of these nonrecurring events. 5. It is possible to analyze naturalistic driving data to characterize typical levels of variability in travel times and to develop measures for quantifying travel time reliability. Limitations of Existing Data Sets The team reviewed multiple naturalistic driving data sets involving video data that are currently available. In analyzing driver behavior, as is the case with this research effort, high-quality video data is required. In general, all existing data sets are satisfactory in terms of video quality because

4driver behavior can be clearly viewed regarding decision errors, performance errors, inattention, and recognition errors. The following limitations still exist: 1. Some studies had fewer video cameras installed compared with other studies. For example, the Automotive Collision Avoidance System (ACAS) FOT and the RDCWS FOT conducted by the University of Michigan Transportation Research Institute (UMTRI) had only two video cam- eras: one facing the driver and the other facing the front view. In these cases, the data sets are limited because traffic conditions beside and behind the subject vehicles are not available. The video frequencies of these UMTRI studies were set relatively low because the original research purposes were not driver-behavior oriented. Consequently, the causal factors of safety-related events are not viewable. 2. Image glare was a typical problem with video data. Some data sets have issues with glare that sometimes make it difficult to make judgments regarding driver behavior. 3. Accidental cable unplugging or malfunction caused incompleteness or errors in data. Although linear interpolation can solve some of the missing data problems, in many cases such problems were not easily detected or corrected. 4. Driver identification is an issue worthy of attention. In a naturalistic driving study, it is not uncommon for the equipped car to be driven by drivers other than the appointed partici- pant. Although the video data can be manually viewed afterward to differentiate drivers in data reduction, it is more efficient if an automatic identification process can be used to tag each trip recorded with driver information so that the data analysis can avoid unnecessary biases. 5. Existing data sources lack a sufficient sample size of crash events to identify changes in driver behavior and the impact of these changes on nonrecurring congestion. The data collection effort in SHRP 2 Safety Project S07, In-Vehicle Driving Behavior Field Study, will offer a unique data set that can be used for further analysis. Recommendations To improve the quality of video data in future data collection efforts of this kind (i.e., designed to investigate the reduction of nonrecurring congestion through modifying driver behavior), there are several recommendations. First, the procedure to recruit participants needs to be carefully designed. It is ideal to include a comprehensive population of drivers ranging evenly across every age, income, and occupation cat- egory. When recruiting participants, it is crucial to make it clear that driver information is vital for the research. To better identify drivers, two methods can be used: 1. A formal statement needs to be included in the contract to make the signer the exclusive driver of the vehicle. 2. A touch-screen device can be installed onboard to collect information before and after each trip. The touch-screen equipment can be designed so that a customized interface will be dis- played to the driver to input trip-related information by selecting certain check boxes. The before trip information-collecting interface may consist of a list of the first names of house- hold members for the driver to select from as passengers, a list of trip purposes, weather con- ditions when the trip started, and any information about why the driver selected the time of departure. The after trip information-collecting interface may include an “original trip pur- pose changed” option, a “route choice changed” option, and a “crashes happened en route” option. Necessary hardware can be designed to connect the input touch screen with the engine so that the driver can start the engine only after the information is input. To ensure safety while driving, the device should be disabled while the vehicle is in motion to prevent driver distrac- tion. One concern with this approach is that it reminds drivers that they are being monitored and thus may deem the study nonnaturalistic.

5Second, to serve the research purpose, certain data are more important than others. The following four categories are imperative: 1. Basic onboard equipment should include devices that collect the following data: video; vehicle network information (speed, brake pedal, throttle, turn signal); global positioning system (GPS) data (latitude, longitude, heading); X, Y, and Z acceleration; distances between the subject and surrounding objects; lane location information (X, Y, Z); driver behavior (seat belt usage, lights on or off); and yaw rate. 2. The video cameras should shoot at least five views: front, back, right, left, and the driver. The resolution should be high enough to identify ongoing traffic conditions, weather conditions, and the driver’s hand movements and facial expressions. Correction of sun glare to improve video quality is available when needed. 3. The frequency setting should be high enough that the video is continuous, the acceleration and deceleration of the vehicles clearly recorded, and the reaction times recorded and measured. The recommended minimum frequency for GPS devices is 1 Hz and for all other equipment, 10 Hz. 4. To improve the versatility of the data so that the data can be used in other related research, vehi- cle performance parameters such as engine speed, throttle position, and torque should be recorded. Third, the data collection system needs to run for an additional 10 minutes after the engine is turned off in case an accident occurs. During the data reduction, data collection usually halted as soon as the driver stopped the vehicle. Because it is important to observe the traffic conditions being affected by a safety-related event, additional data are required after a driver turns off the engine. One concern is that if some malfunction to the subject vehicle occurs (in case of an accident), gathering data may cause a safety hazard. This issue needs further investigation. Fourth, to improve linking vehicle data with external data, it is ideal to standardize the format for time and location information. For vehicle data, the synchronized GPS clock should be used rather than local computer time for better connection of the data with external traffic, crash, work zone, and weather data. For external data, some states have their database built on the milepost sys- tem. The conversion of mileage post locations to a standard latitude and longitude should be con- ducted ahead of time. Fifth, because a limited number of crashes—especially severe accidents that affected traffic conditions—occurred in all the candidate data sets, certain adjustments are needed to create a sta- tistically significant database. A longer data collection effort or more drivers involved in the study would be ideal. For example, SHRP 2 Safety Project S07, In-Vehicle Driving Behavior Field Study (a 2,500-Car Study), which will soon be conducted, is a quality candidate. Another solution is simulation, which can be used to compensate for data shortage. Sixth, additional analysis of existing data is required to study typical levels of variability in driver departure times and in trip travel times and the level of variability in driver route choices. A char- acterization of this behavior is critical in attempting to quantify and develop travel time reliability measures and to understand the causes of observed travel time reliability. The data may be augmented with tests on a driving simulator to study the impact of travel time reliability on driver route-choice behavior. Finally, although numerous studies have used video cameras to gather data, an ideal starting point is a compiled data source list that summarizes existing video-involved studies with specifi- cations of data collected, limitations of data usage, and access issues. Such a list will help prevent redundancy in future investigation efforts.