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.
35 C h a p t e r 8 The emphasis of this project has been on creating and demon- strating feasible analysis methods for the SHRP 2 naturalistic driving study, using limited data that already exists. While the main thrusts of this analysis have been in Bayesian estimation and extreme value theory, the research questions posed can be addressed by other techniques. In this chapter the research team proposes simple additional relevant analysis techniques that can be used to resolve the data in different (orthogonal) directions. The first of these techniques considers how exist- ing GIS tools for the analysis of spatial clustering may be directly applied to road departure crash data, with the aim of suggesting crash patterns and hence formulating hypotheses for deeper analysis. The second technique relates to driver distraction and surrogacy by comparing surrogate event rates in episodes of driving while on and off a cell phone. The third technique compares a measure of control disturbance (YRE) in the presence of particular highway features, based on the tentative hypothesis that disturbed control is more likely when lane or road boundaries change, for example, when the road widens near the entrance to a side-street. As with the main analysis approaches that have been used, the aim is to propose and test viability for application to the future SHRP 2 driving database and not to answer specific research questions based on the limited data used. road Departure Crash Spatial analysis Examining common features of spatial clusters of crashes, or hot spots, can be useful in identifying patterns in the occur- rence of such events. The common features of the hot spots can be related to roadway characteristics or to driver behav- iors captured in NDD, or to both, and can be informative in developing hypotheses for further analyses. For example, examination of roadway data from the hot spots may suggest roadway features associated with the crashes and may pro- vide a nonparametric indicator of the existence of surrogates. Examination of driving data at the hot spots may suggest pat- terns or particular driver behaviors for deeper analyses. What appear to be concentrations of crashes, however, could simply be a result of random occurrences, and looking for commonalities in these random groupings would be a wasted effort. Spatial analysis tools allow researchers to identify clus- ters that constitute statistically nonrandom, spatially depen- dent hot spots. In this orthogonal study, ArcToolbox spatial analysis tools are used to identify nonrandom spatial clusters of run-off-road crashes. The development of comprehensive information about the hot spot clusters through the joining of spatially referenced crash locations, roadway, and roadway characteristics is also demonstrated. In spatial analysis, hot spots are formally defined as clus- ters of points with values higher in magnitude than one would expect to find by random chance. The spatial analysis of 7,924 road departure crashes in Oakland County, Michigan, from 2001 to 2005 was completed by using the Getis-Ord Gi* sta- tistic. The Getis-Ord Gi* statistic is a well-known method for establishing local spatial dependencies that may exist and one of the six statistics designed to address local relationships among geo-referenced data (Getis and Ord 1996). The conceptualization of spatial relationships used for weighting the crash data was defined by using the inverse and Euclidean distances. The threshold distance chosen for the calculation was zero. The inverse distance method gives more weight to those crash locations that are close to each other and less weight to those farther apart. The Euclidean distance models the shortest path between events, and a threshold dis- tance of 0 ensures that no points are dropped from the analy- sis. The road segment crash rate was selected as the random/ independent variable or our exposure value. Given a set of weighted data points, the model generated a Z-score for each segment along which a crash occurred. Hot spots are indicated by a Z-score greater than +1.96 or locations with a high crash-rate cluster. The hot spot location has a high crash-rate value and its neighbors also have high crash-rate Orthogonal Studies
36 reveal some patterns worthy of further pursuit. However, it was not the intention of this exploratory study to establish common factors of the crashes or go into further analysis, but to demonstrate the use of spatial analysis tools and to suggest that such an investigation would be fruitful. The results of any such microscopic analysis may suggest possible common ele- ments and help formulate a hypothesis for more formal study. Cell phone Use Engaging in secondary tasks while driving taxes a driverâs attention resources, leaving less for the driving task. In many normal driving situations the attention demand for driv- ing is met even with secondary tasks without crashes or other incidents. However, whenever attention demand exceeds available attention resources, the risk of a crash increases (Eby and Kostyniuk 2004a). Carrying on a cell phone conversation while driving has been shown to negatively affect driving values or higher than would be expected by random chance. These hot spots indicate nonrandom, spatially dependent crash-rate locations. Running the model resulted in only a few (397) of the 7,924 crashes having a Z-score greater than +1.96 or an indi- cation of spatial dependency. There were no Z-score values less than or equal to -1.96, which means that no âcold spotsâ were found. Cold spots are locations where low crash rates are surrounded by locations with similarly low crash rates. The orange dots on the top of Figure 8.1 indicate the hot spot locations. The aerial image zooms in on a hot spot cluster. The crashes, digital base map segments (the blue line), and signage can be viewed as a layer on the aerial photograph. The data table illustrates the selected HPMS road segment data that were associated with crashes. Although this hot spot is located near an intersection, the crashes considered here are single-vehicle run-off-road crashes. Further analyses of the crash data or of driving data might Sources: Google Earth, ArcMap XY plot using project data, and Uniform Traffic Control Manual. Figure 8.1. Clusters of run-off-road crashes, Oakland County, Michigan.
37 it. A different analysis of secondary tasks in the same RDCW data by Sayer, Denvonshire, and Flannagan (2005) concluded that drivers select the time and conditions for their cell phone conversations and that there is a large amount of variability in the behavior between drivers, and even of the same driver on different days. Studies with tight experimental design and controls will be challenging at best with natural use driving data. A different approach might be more suitable for analyzing the effect of cell phone use (or other secondary behaviors) on driver per- formance in natural use data. Rather than designing a study around episodes of cell phone use, one could design a study around the candidate marker of disturbed control (i.e., YRE or LDW), identify events or conditions in which critical values of these markers occur, and then examine the environmental and driver conditions that were present. road Boundary Change and Yaw rate error Chapter 7 introduced the YRE estimation and explained how it might be used to determine the lane-tracking performance of drivers. This short study investigated the possibility that a change in road geometry, in the form of a deviation in the perceived road boundary, would influence the lane-tracking performance. The YRE was calculated for multiple passes past a point where the road provides deceleration/acceleration lanes (or generous turning radii) at T-intersections and com- pared that to the YRE for travel with uniform boundaries. Ten intersections were selected for analysis based on a ran- dom sampling of satellite images, and of these, two were tra- versed multiple times by various drivers in the NDD. Figure 8.2 shows satellite images of the two points used in the analysis. Notice that both points display a road widening situation for only one direction of travel, which enabled the traversals in the opposite direction to be used as a baseline for compari- son. Point 1 had 35 traversals by four unique drivers, 23 in the direction with the widening and 12 that could be used for a baseline. Point 2 had 95 traversals by 10 unique drivers, 56 in the direction with the widening and 39 baseline passes in the opposite direction. The YRE was calculated for 5 s before and after the middle of the intersection for each pass in both directions. Figure 8.3 shows the standard deviation of the YRE for each 10-s pass. Even without formal statistical analysis, in each case there appear to be no systematic differences between the underlying samples shown in red and blue. This is not particularly sur- prising, given the small size of the samples. The implication is that the general driving pattern is unaffected by the road boundary change. On the other hand, the break, or devia- tion in the perceived road boundary, may occasionally pro- vide an outlier caused by a corresponding disturbance in the performance in many simulator studies (Caird et al. 2004; Hoerrey and Wickens 2004). There is also much anecdotal evidence of unsafe driving behavior of drivers on cell phones. This suggests that the effectiveness of tactical and operational aspects of the driving task is reduced when a driver is engaged in a cell phone conversation. Furthermore, if this is true, then there should be some evidence of disturbed control in the vehicle kinematic history of the driving episode. This study explored LDW alerts as a marker for disturbed vehicle control in a comparison driving with and without a cell phone. Episodes of cell phone use were identified in the RDCW naturalistic driving database by visual review of face videos of the first week of driving for all subjects. Compari- son episode periods were selected for each cell phone episode of each driver randomly from the same trip and on the same road type. On average, each subject in the RDCW study drove the vehicle on 32 trips in the first week of driving. Of the 78 sub- jects, 61 used the cell phone in the first week, spending an average of 1,150 s on that activity. The number of cell phone episodes per driver ranged from 1 to 64, with an average of 11.6 episodes. The duration of the cell phone episodes ranged from 4 to 818 s, with an average of 99 s. Of the 61 subjects, 25 (41%) did not trigger an LDW alert while on a cell phone or during the comparison periods. Of the 36 drivers with LDW alerts, 18 (50%) triggered alerts only in comparison periods (i.e., while not on a cell phone), 4 (11%) triggered alerts only during cell phone episodes, and 14 (39%) triggered alerts in both cell phone and comparison episodes. The overall LDW alert rate for each driver was developed from the cell phone usage on all of his or her trips, and the comparison rate was developed from all the comparison peri- ods. Of the 14 drivers with LDW alerts in both the cell phone episodes and in the comparison periods, the rate of alerts during cell phone periods was greater than the rate for the comparison period for eight drivers and less than the rate for six drivers. The exploration of LDW alerts and cell phone does not show a greater rate of LDW alerts while drivers are on cell phones. Indeed the indications are that in general, drivers are more likely to receive an LDW alert while not using a cell phone than while on a cell phone. Because YRE appears to be a useful metric for lane excur- sions, the team examined its applicability to this situation also. YRE was calculated for episodes of cell phone use and comparison periods for several subjects. This exercise also did not show much difference between driving periods with cell phone and without cell phone usage. While it seems that there should be some indication of a deterioration of vehicle control while the driver is engaged in a cell phone conversation, this approach was not able to identify
38 (a) (b) Source: Google Earth. Figure 8.2. Road widening on right boundary for (a) Analysis Point 1 (42.276308, â83.662979) and (b) Analysis Point 2 (42.320524, â83.539831). (a) Point 1 Driver Pass 0 5 10 15 20 0 0.002 0.004 0.006 0.008 0.01 0.012 0.014 0.016 0.018 0.02 st .d ev . y re [ra d/s ] Widening Boundary Uniform Boundary (b) Point 2 Driver Pass 0 10 20 30 40 50 0 0.002 0.004 0.006 0.008 0.01 0.012 0.014 0.016 0.018 0.02 st .d ev . y re [ra d/s ] Widening Boundary Uniform Boundary Figure 8.3. Standard deviation of the estimated YRE for travel in the direction of the broken boundary and complete boundary (baseline driving conditions) for (a) Analysis Point 1 and (b) Analysis Point 2.
39 While the directional differences are small and inconclusive, differences in location are very clear. Comparing Point 1 and Point 2 for the mean levels of YRE (aggregated over both direc- tions), for Point 1 there is a mean ± one standard deviation of 0.0078 ± 0.0035 rad/s, while this measure for Point 2 (both directions) is 0.0035 ± 0.0022 rad/s. The YRE calculation dem- onstrates sensitivity to the differences in lane-tracking perfor- mance between different locations. Clearly there is a difference between the driving that occurred on Points 1 and 2. Figure 8.3 shows that the driving near Point 1 had a larger variation in YRE than did the driving near Point 2, indicating less uniform lane tracking and more driver corrections for Point 1. The rea- sons behind this difference have not so far been investigated. lateral control. The fact that the blue outliers here are the more extreme is somewhat supportive of this idea, but clearly more data would be needed to provide definitive results. YRE is pre- ferred to simpler lane-tracking measures, given its apparent strength of connection to control disturbances, and extreme values might be more productive than using standard devia- tions. To gather more data, a pool of corresponding sites could be used for aggregation, or it might be that in the future large- scale driving study, a larger sample of relevant traversals could be found. The point here is that a study relating specific safety- related driver behavior to highway features appears feasible, but that a credible connection to disturbance in control is an essential ingredient.
