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.
48 C h a p t e r 1 0 In Chapter 2, three general research questions were stated, together with 14 more specific questions that were intended to guide this research. The aim, even in the most optimistic moments, was not to answer all questions in a single project but to develop and test methods that have a strong bearing on these questions. The broad research questions were: ⢠Do single-vehicle road departure crashes occur only under conditions of disturbed control? ⢠Do naturalistic driving data contain measurable episodes of disturbed control? ⢠Do objective measures of disturbed control from natural- istic driving data, together with highway geometric factors, off-highway factors, and environmental factors, satisfy criteria for crash surrogate (i.e., are they related to actual crashes)? On the one hand, the SUR analysis and the analysis of YRE in Chapter 7 provide ample indication that episodes of disturbed control exist in the naturalistic driving data, and that these episodes can be related to crashes via high- way variables. On the other hand, insufficient data exist for the direct inclusion of off-highway factors, and little or no investigation was made in the area of other environmental factors such as lighting and weather conditions. From the data analyzed, it has not been possible to make any detailed progress on the first of the broad research questions: Do single-vehicle road departure crashes occur only under con- ditions of disturbed control? This lack of progress is not sur- prising, given that the data do not contain actual crashes. For the future SHRP 2 vehicle study, it should be a priority to estab- lish this association. Given the very approximate estimation from Chapter 6 that around 15% of road departures lead to actual crashes (at least on the road type considered), it should also be a priority to capture episodes of drift off the paved highway. In this way the preconditions for an expanded set of events might be used to analyze the problem, increasing the number of events and removing the factors that determining whether an actual impact occurs. Turning next to the more specific research questions, sig- nificant progress has been made on the following research questions. Question 1: What measures exist in naturalistic driving data that directly measure disturbed control? The candidate surrogates LDW and TTEC provide some general measure of disturbed control, as response-only variables, while YRE appears to provide a more focused and control-related measure of this aspect. Question 2: Are vehicle kinematic measures sufficient to identify disturbed control for risk measures in single-vehicle road departure crashes? It has been shown that vehicle kinematics can provide a basis for risk analysis and surrogate validation. Whether this can be improved by using direct measurement of driver action has not, however, been analyzed. Question 3: Are other driving control metrics necessary (in addition to vehicle kinematic measures) to identify dis- turbed control? This question has not been resolved within the current data and time limitations of the project, but it is surely feasible in terms of some of the surrogates formulated but not analyzed in Chapter 4 (especially those relating to coherency or visual interruption). Question 4: Are there measures of driving control perfor- mance in existing FOT data that depend on highway factors in a way that is consistent with single-vehicle road departure crash frequencies? The prime success of the research presented relates to this research question. Through SUR analysis, it is certainly the case that measures of control performance can be related to crash data via the use of independent highway factors. Question 9: What statistical tests are available to determine if the measures of driving control performance in naturalistic Conclusions
49 Question 12: Can various driver states (e.g., drowsy, aggres- sive, distracted, engaged) be identified from naturalistic driv- ing data? Question 13: Can driving control performance for various states be categorized more simply (i.e., good and bad, or risky and nonrisky)? Question 14: Is there a difference in the driving control performance of good and bad drivers (or risky and nonrisky drivers) at locations with geometric features associated with high single-vehicle crash frequency? Some progress has been hinted in analysis of the YRE crite- rion. In Chapter 7, the figures presented indicate that a coher- ent interaction between yaw rate, yaw rate critical limits, and steering intervention is at the heart of active and effective lane-keeping control. Further work on this is necessary, but if successful, it may make active monitoring of driver gaze less critical to the assessment of driver attention state during driving. The cell phone study described in Chapter 8 under Cell Phone Use was unable to show any systematic differ- ences between with cell phone and without cell phone, but a number of confounding factors may exist (especially in terms of the choice of time to speak on the phone). Again, a more sensitive measure of lane keepingâone based around coherency rather than simply on error magnitudesâmay again be helpful. Thus there are suggestions that questions on detection and categorization of driver state (Questions 12 and 13) may be achievable. It should be noted, however, that this study relied on lateral control measures, and it is certainly the case that further analysis based on combined longitudinal and lateral control would be more complete in addressing those two questions. In this study, no analysis was considered in relation to interactions between driver types and highway risk features (Question 14). In the future it is quite possible to include such factors in the SUR analysis. Turning to other aspects of the study, in Chapter 2 under Data Quality and Validation, the research team posed a number of questions about data availability and quality. According to Chapter 9, it is shown that the data types and quality are not unique to Michigan data. The Virginia DOT maintains spatially referenced data for crash and highway factors, and the SHRP 2 naturalistic driving study will pro- vide many of the variable contained in the UMTRI NDD, expanded in some areas (such as head pose estimation) though reduced in other areas (such as number and location of radar units). Most, if not all, of the analysis conducted here with Michigan data is also feasible with Virginia data, though resources were too limited to engage in full replica- tion studies. In the SUR analysis it was recognized that a novel approach should be taken to exposure, with event rates based on seg- ment traversals. In refining this idea, directional segments were defined, but consideration of curves made it clear that data and single-vehicle crashes depend on geometric features in a consistent way? SUR analysis provides a partial answer to the question of what statistical tests are available. Question 10: Can satisfactory crash risk predictions be made on the basis of vehicle/driver/highway information from natu- ralistic driving (e.g., via extreme value theory), or do additional roadside and environmental factors need to be introduced? Extreme value theory suggests that a positive answer to Question 10 is also feasible, though in this case only a sim- ple initial analysis was made. However, the potential benefits of introducing other roadside and environmental factors (in some form of composite surrogate) have not been tested. Little progress has been made on the following research questions. Question 5: Are there specific highway features that are associated with single-vehicle road departure crashes and specific driving control performance measures? A small example of data exploration was presented for cases in which the right lane boundary was poorly defined or ambiguous, but it was not possible to find large numbers of such locations. While the optical lane tracker has some potential to be used in this way, a small attempt to conduct such an analysis did not prove fruitful. The lane tracker, when used to track broken or missing boundaries in this way, sim- ply did not create reliable results (checking video revealed no useful patternsâi.e., the lane tracker detection states for the UMTRI NDD were not reliable predictors of road marking and lane conditions). Question 6: Can roadside factors (e.g., locations of poles, trees, bridge abutments, and side slopes) be coupled to natu- ralistic driving data? Question 7: Does the coupling of roadside factors to naturalistic driving data improve correlation with actual crashes? Question 8: Can general descriptors of roadside environ- ments be used in this coupling (e.g., tree density, proportion of side slope steeper than 4 to 1), or do we have to be more specific about location of roadside obstacles? For Questions 6 to 8, relating to roadside factors, progress was not possible because of the lack of objective data for fea- tures such as cross-slope, trees, and poles. This lack of data was the most disappointing aspect of the study, especially since it is quite possible that SUR analysis could be improved with these additional factors used in the crash model (though naturally excluded in the surrogate model). There was again little objective data directly available relat- ing to driver state for the questions concerned with driver factors. Question 11: Is the pattern of driving control performance different for the same driver when distracted versus not dis- tracted (e.g., on a cell phone or not on a cell phone)?
50 A final conclusion relates to the orthogonal studies pre- sented in Chapter 8 and also the data analysis presented in Chapter 9: the combination of naturalistic, crash, and high- way data provides such a rich data resource for research that it is relatively easy to develop new and innovative analysis methods specific to the data types. Therefore, making these data types available to independent researchers in the future seems certain to spawn new ideas and gather new insights. segment definitions were in need of further revision. There is a basic need for homogeneity in the segment with respect to the highway factors of interest. As such, HPMS segmenta- tion is not generally appropriate, so in the future it makes good sense for safety researchers to define their own segment boundaries. Again, time was not sufficient to carry this analy- sis through, but it is clearly an important insight into how best to conduct future SUR analysis.
