Accident Analysis and Prevention 74 (2014) 1–7
Contents lists available at ScienceDirect
Accident Analysis and Prevention journal homepage: www.elsevier.com/locate/aap
Characteristics of turn signal use at intersections in baseline naturalistic driving John M. Sullivan a, *, Shan Bao a , Roy Goudy b , Heather Konet b a b
University of Michigan Transportation Research Institute, 2901 Baxter Rd., Ann Arbor, MI 48109-2150, USA Nissan Technical Center North America, 39001 Sunrise Drive, Farmington Hills, MI 48331-3287, USA
A R T I C L E I N F O
A B S T R A C T
Article history: Received 25 March 2014 Received in revised form 2 October 2014 Accepted 6 October 2014 Available online xxx
The purpose of this study was to determine whether a driver’s use of turn signals is sufficiently reliable to forecast a vehicle’s future path around an intersection, when detailed information about the intersection is unavailable. Naturalistic observations of turn signal use among 108 drivers on surface streets were extracted from the baseline portion of a field operational test of a safety system. Left and right turns that resulted in heading changes of between 70 and 110 and turn radii between 18 and 90 m were selected from the dataset. The odds that a driver would signal a turn were modeled as a function of road type, turn direction, presence of a forward vehicle, whether the vehicle stopped before the turn, and driver age and gender. Overall, 25 percent of left turns and 29 percent of right turns were not signaled. Road type, turn direction, and presence of a forward vehicle were found to influence the odds that a turn is signaled, while gender and age of the driver did not. The results suggest that situational factors like road type and turn direction are more powerful predictors of whether a turn will be signaled than either age or gender. Signaling on major and minor surface roads was about 5 times more likely than on local roads and 1.5 times more likely when a forward vehicle was present, suggesting a possible effect of traffic volume. It was concluded that turn signal activation alone may be insufficiently reliable to forecast a driver’s path. ã 2014 Elsevier Ltd. All rights reserved.
Keywords: Turn signal use Surface streets Turn direction Forward vehicle
1. Introduction In this study, naturalistic data were examined to help determine what systematic characteristics can be observed in a driver’s turning maneuvers that might be used to inform the development of a connected-vehicle-based intersection collision avoidance (ICA) application. Such an application would warn the driver not to proceed along a projected course if there is a high likelihood that a collision with another vehicle might occur. For such a system to do this well, it must be able to anticipate what the driver intends to do as the intersection is approached as well as determine what the other vehicles in the immediate vicinity are also likely to do. Prior attempts to address this type of road conflict have experimented with both vehicle-based radar to assess conflict (e.g., Pierowicz et al., 2000) as well as site-based radar approaches (Misener et al., 2010; Nowakowski, 2006; Penney, 1999). With the new capabilities enabled by connected vehicle technologies (also known as
* Corresponding author. Tel.: +1 734 764 8560; fax: +1 734 764 1221. E-mail addresses: [email protected] (J.M. Sullivan), [email protected] (S. Bao), [email protected] (R. Goudy), [email protected] (H. Konet). http://dx.doi.org/10.1016/j.aap.2014.10.005 0001-4575/ ã 2014 Elsevier Ltd. All rights reserved.
dedicated short range communications – DSRC), equipped vehicles within an approximately 300-m radius of each other will be able to transmit to and receive detailed information from other nearby vehicles. This form of wireless communication has been called vehicle-to-vehicle (V2V) communication to distinguish it from other kinds of connected-vehicle communications involving interactions between the vehicle and the local infrastructure (called V2I). Advantages of a V2V communications capability were noted more than 10 years ago by Miller and Huang, 2002 who proposed a peer-to-peer data sharing algorithm to be used to address intersection collision crashes with possible extensions to address both frontal and rear-end collisions. In their framework, the peer vehicles primarily shared position data derived from global positioning systems (GPS). Route contention was determined by calculating trajectories from the GPS data, and warnings were issued if the time-to-collision (TTC) fell to a value near to the time a driver was projected to require in order to avoid the collision. This time was called the time-to-avoidance (TTA). If the driver took any mitigating action (e.g., braking) before TTC became too low, the warning was withheld. Unlike an exclusively GPS-based system, V2V transmissions contain additional information that describes driver actions in nearby vehicles. Besides latitude, longitude, heading, and speed
2
J.M. Sullivan et al. / Accident Analysis and Prevention 74 (2014) 1–7
(available from on-board GPS), the basic safety message may also contain information about steering angle, braking, and turn signal use (see Society of Automotive Engineers, 2010 for message details). Thus, a collision avoidance system may be able to better predict a driver’s intention to stop, turn, or proceed through an intersection based on some of this additional information. A warning system’s reliance on this information will likely depend on how well it predicts what a driver will do in the immediate future. For example, recent work on turn signal use reports systematic variation in drivers’ use of turn signals. In two observational studies, Faw (2013) observed the turning maneuvers of drivers of passenger vehicles tabulating whether there was no signal, a late signal, or a good signal. In the first study, 3149 vehicles were observed executing both left and right turns at 47 different intersections; in the second study, selected based on low rates of turn signal use, 2455 right turns were observed at two intersections. Signaling was performed less often when turning right versus turning left, and less often when a dedicated turn lane was present. Turns at intersections were reported to be signaled about 76 percent of the time. This suggests that a turn is likely to be executed when a driver is signaling, but a turn may also be made 24 percent of the time when the driver is not signaling. Presence of traffic was also found to influence use of turn signals: in one study (Lebbon et al., 2007), 63 percent of drivers signaled turns in the presence of oncoming traffic compared to 44 percent when no traffic was present. On the other hand, Faw (2013) reports that turn signal use is lowest in heavy traffic (79 percent signaling), compared to light (89 percent) and moderate traffic (92 percent), suggesting less signaling in the presence of other traffic. The discrepancy might be a consequence of the small number of intersections sampled – Lebbon et al. (2007) sampled drivers at two intersections; Faw (2013) sampled drivers at 22 different intersections. Intersection characteristics and road configuration have also been noted to influence driver use of turn signals. For example, Faw (2013) reports 87 percent signaling when a dedicated turn lane is present, compared to 90 percent when no lane is present. The purpose of the present analysis is to examine naturalistic vehicle turning maneuvers for systematic regularities that could help indicate a driver's intention to turn as the driver approaches an intersection. This study examines turn signal use and its association with driver demographics (age and gender), and other factors derived from the naturalistic driving dataset including road type, turn direction, and the presence of lead vehicles on the roadway. Some of these factors could also be considered reasonable surrogates for traffic density – for example, it might be expected that arterial roads have a higher traffic density than local roads; likewise, the presence of a lead vehicle may be indicative higher traffic density than no lead vehicle. We also note that the naturalistic data differs in significant ways from the observational studies: driver age and gender are known; the drivers involved are all driving the same make and model vehicle (a 2006/2007 Honda Accord); and the set of 3830 turns in the sample were drawn from approximately 2732 unique locations. 2. Methods 2.1. Naturalistic data sample
(older), resulting in 18 drivers in each cell. Drivers were given test vehicles to drive unsupervised for a 12-day baseline period in which safety systems were inactive, followed by a 28-day period in which the safety systems were activated. The sample used for these analyses were derived from the baseline period of driving, when the safety systems were inactive. Thus, the baseline represents driving a vehicle equipped with no special warning systems that might artificially raise a driver’s awareness of the need to signal turns. 2.2. Turn selection criteria 2.2.1. Road type Because the targeted countermeasure scenario is intersection collision avoidance, selection of turn maneuvers was restricted to surface streets. This eliminated ramps, limited access roadways, and unknown road types which are largely parking lot maneuvers. Three road types were included in the analysis: major surface, minor surface, and local roads. Road types were derived from a commercial mapping database’s (NavTeq/HERE) five function classes and whether the roadway is identified as a limited access roadway or ramp. Major surface streets included roads supporting moderate-speed travel within cities and travel between cities (function class 2 and 3); minor surface streets included moderatespeed travel between neighborhoods (function class 4); and local roads were defined as supporting lower-speed travel between neighborhoods (function class 5). 2.2.2. Turn characteristics Turn maneuvers were selected that involved a heading change of between 70 and 110 . Turns were initially detected by examining vehicle yaw rates that exceeded 8 s 1. The beginning and end of the turn was determined by backtracking and forward-tracking in the yaw-rate time series until the absolute yaw rate dropped to about 0.5 s 1. This produced an approximate start and end of a turn (see the example turn shown in Fig. 1). Change in heading was determined by calculating the difference in heading at the start of the turn and the end of the turn. A negative difference was used to indicate a left turn; a positive difference indicated a right turn. After direct video inspection, several turns in the initial sample were found to occur in parking-lots or parking structures and were unrelated to intersection traversal; this occurred because the proximity of the parking area to a nearby road resulted in the
-83.4432 -83.4433 -83.4434 -83.4435 -83.4436 -83.4437 -83.4438 -83.4439
Turning data were developed from the driving data obtained in the Integrated Vehicle Based Safety System naturalistic driving study (Sayer et al., 2011). In this study, 108 licensed drivers were recruited to drive passenger vehicles equipped with a combination of safety systems. The sample of drivers was equally divided between male and female drivers and stratified into three age groups: 20–30 years (younger), 40–50 (middle-aged), and 60–70
-83.444 42.5449 42.545 42.545142.545242.545342.545442.545542.545642.545742.5458
Fig. 1. GPS trace of path of a vehicle executing an approximately 90 turn; the red circles identify the part of the turn examined in the analyses. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)
J.M. Sullivan et al. / Accident Analysis and Prevention 74 (2014) 1–7
3
Fig. 2. Frequency distribution of turn radii among the selected curves. The red vertical line is the mean of each distribution. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)
vehicle being erroneously located on one of the acceptable road classes. In addition, some 70–110 changes in direction also occurred on uninterrupted roadways and did not involve intersection maneuvers at all. This usually occurred along rural roadways. To address these issues, turns were further filtered by determining whether the turn occurred within 5 m of an identified roadway node, using State of Michigan roadway geometric data derived from the Center for Geographic Information (http://www. michigan.gov/cgi). One consequence of this filtering was that driving that occurred outside the State of Michigan was excluded from the sample. Turns were also filtered based on average turn radius, calculated from the total distance traveled over the turn and
the total heading change. Turns with a radius less than 18 m were most often observed in parking areas and residential driveways; turns with a radius of more than 90 m were most often observed on continuous rural roadways and not related to intersections. These criteria were initially established based on the Delaware State Department of Transportation design criteria for the inside radius of intersections having a design speed of between 24 and 56 km/h (15–35 mph) (Delaware Department of Transportation, 2006). Direct video inspection was used to confirm this selection strategy. Following application of these filtering criteria, a pool of 3830 turns remained in the data sample, of which 2732 occurred at unique locations.
Fig. 3. Length of turn by direction and road type. The red vertical line identifies the mean of each distribution. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)
4
J.M. Sullivan et al. / Accident Analysis and Prevention 74 (2014) 1–7
Fig. 4. Duration of turn by direction and road type. The red vertical line identifies the mean of each distribution. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)
3. Results 3.1. Characteristics of sample turn paths The sample of turns was characterized with respect to distribution of continuous variables related to turn execution – turn radius, turn length, turn duration, heading change – over categorical factors that included turn direction, road type, driver age group, and gender. Each of the continuous variables was modeled as a dependent measure using the categorical factors as predictors in a mixed model analysis that included driver as a random factor. The statistics reported here describe the Bonferroni-adjusted pairwise comparisons between factors in which systematic differences were observed for each dependent measure. In general, the turn radii of local roads were systematically shorter by 2–3 m than either major or minor surface streets (t = 3.3, p < 0.01, and t = 5.0, p < 0.01, respectively; see Fig. 2), and the turn radii of right turns was shorter than left turns by about 1.6 m (t = 4.2, p < 0.01). These radii were based on the actual turn paths that drivers followed while executing turns and do not reflect the geometric design of the roadway. No effect of driver age or gender was observed. Consistent with the turn radii data, the distribution of turn length was also systematically shorter on local roadways than major and minor surface streets and shorter for right turns than left turns (shown in Fig. 3). The turn length along local roads was observed to be an average of 3–5 m shorter than major or minor surface roads (t = 3.1, p < 0.001, p < 0.001, t = 5.4, p < 0.001, respectively). The average length of right turns was about 3 m shorter than left turns (t = 4.8, p < 0.001). No effects of driver age or gender were observed. The duration of turn execution averaged between 7 and 8 s, taking about 0.4–0.5 s longer from local roads than from major or minor surface streets (t = 5.5, p < 0.01; t = 5.6, p < 0.01, respectively). No systematic differences were observed in turn duration between left and right turns (see Fig. 4). Younger drivers traversed turns in about 0.45 s less time than middle aged and older drivers (t = 4.0, p < 0.01; t = 4.1, p < 0.01).
The distribution of heading changes after selection of turns conformed to the approximate 90 left and right heading changes expected after application of the turn selection strategy. No systematic differences were observed in the sample heading changes were across the factors (Fig. 5). 3.2. Sample driver distribution While care was taken to ensure that there was a relatively balanced distribution of male and female drivers of each age group that participated in the IVBSS study, the study did not regulate drivers’ exposure to turns. Drivers who logged more hours driving were likely to execute turn maneuvers more often than drivers who did not. The proportional makeup of the driver sample is shown in Table 1. The sample of turns has a greater proportion of young drivers than middle aged and older drivers. Within the young and middle aged categories, proportionally more turns were executed by male drivers than female drivers; the gender split between the older drivers is approximately even. The distribution of gender and age is not independent (x2 = 17.8, p < 0.001). The generalized linear mixed model analyses presented later includes driver as a random factor to ensure that results are not biased by disproportionate numbers of observations by driver. The distribution of driver age across turn direction was also examined in light of prior work that suggests older drivers may self-limit their driving (Braitman and Williams, 2011), perhaps to avoid complicated roadway situations. The sample was examined for evidence of any particular age-related selectivity in turning maneuvers. As can be seen in Table 2, no evidence was found to suggest that turn distributions among older drivers differed from younger and middle aged drivers (x2 = 0.11, p = 0.94). 3.3. Driver’s use of turn signals For each turn in the dataset, turn signal use was identified for all cases in which the turn signal was actuated at some point within the turn execution. For most turns, activation occurred before the turn began, although, drivers occasionally began signaling after the start of the turn. For the present analysis, both kinds of signal
J.M. Sullivan et al. / Accident Analysis and Prevention 74 (2014) 1–7
5
Fig. 5. Heading change distribution by direction and road type. The red vertical line identifies the mean of the heading change distribution. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)
vehicle. This eliminated most stationary targets, vehicles traveling toward the subject vehicle, and vehicles moving away from the subject vehicle at more than twice the subject vehicle’s current speed. The intention of this selection strategy was to identify situations in which the driver’s approach to an intersection might be constrained by the presence of other vehicles in proximity to the subject vehicle. Secondarily, the indicator identifies situations in which forward traffic is present as the intersection is approached. Vehicles that stopped before the turn were identified as those whose speed was less than 1 m s 1 any time in the 100-m distance before the start of the turn. Vehicles commonly stop in response to traffic signals or other traffic impediments like cross-intersection traffic.
actuation are considered as signaling. The breakdown of turn signal use in the study sample is shown in Table 3. Overall, unsignaled turns occurred in about 25 percent of the left turns, and 29 percent of the right turns. Signaling errors, in which a turn is made in the direction opposite to the direction of the turn executed, were quite rare – about 0.1 percent. One of these cases resulted from a left turn off a right-curving road that was misidentified by the analysis algorithm. The other two cases involved one driver who appeared to abruptly change their mind about the turn direction (e.g., their gaze was initially toward the signaled direction but abruptly changed), and another who signaled to exit a left-turn lane, abandoned the attempt, and turned left. Generalized linear mixed models were used for this part of the analysis and account for repeated observations of the same driver. The between-subject variables were age (young, middle, old) and gender (male, female). The four within-subject variables were roadway type (local, major surface, minor surface), direction of the turn (left, right), presence of a forward vehicle moving in the same direction as the subject vehicle within a distance of 100 m (yes, no), and whether the turn was preceded by a stop (yes, no). In the mixed model, driver and interactions between the driver and within-subject effects were treated as random effects. This accounts for within-subject correlated observations and effectively allows comparison of a driver to him/herself. The dependent variable in this analysis was whether a driver signaled the turn. Presence of a forward vehicle was determined by scanning the vehicle’s approach to the start of the turn (beginning 100 m back) for the presence of any targets in the forward radar within 100 m and traveling at a relative absolute speed less than the subject
3.3.1. Analysis results Three models were constructed and their fit statistics compared. An intercept-only model was used to establish a base model with none of the predictors present. A second model was constructed that included all of the predictors. And finally, a third model was evaluated that included only predictors in which a significant association with the dependent measure was observed. The fit statistics of each model are shown in Table 4 for comparison. The results suggest that the fit of the model that includes the significant predictors is best, and is the one selected for discussion. Main effects were observed for turn direction, road type, and forward vehicle presence. No effects were observed for driver age group or gender. A marginally significant effect was also observed in situations in which a vehicle stopped before initiating the turn. Odds ratio estimates are shown in Table 5. Left turns were about 1.38 times as likely to be signaled as were right turns. There was
Table 1 Distribution of sample curves among drivers in sample.
Table 2 Distribution of sample turns by driver age and turn direction.
Age group
Age group
Gender
Young
Middle
Older
Total
Turn direction
Young
Middle
Older
Total
Male Female Total
906 637 1543
646 506 1152
574 561 1135
2126 1704 3830
Left Right Total
746 797 1543
557 595 1152
542 593 1135
1845 1985 3830
6
J.M. Sullivan et al. / Accident Analysis and Prevention 74 (2014) 1–7
Table 3 Turn signal use by turn direction. Turn direction
Counts
Signal status
Counts
Percent
Left
1845
Unsignaled Signaled left Signaled right Unsignaled Signaled left Signaled right
453 1390 2 570 1 1414
24.6 75.3 0.1 28.7 0.1 71.2
Right
Total turns
1985
3830
little difference in the odds of signaling between major and minor surface streets, however each differed from the odds of signaling on local streets by more than 4 times. When a forward vehicle was present, signaling was about 1.5 times as likely as it was when no forward vehicle was present. Finally, signaling was also about 1.3 times as likely if a vehicle had stopped in the 100 m preceding the turn. 4. Discussion One interpretation of the observed results is that they may be related to the likelihood that other vehicles are present on the roadway to motivate the driver to communicate the intended maneuver. On sparsely populated roadways, drivers may be less inclined to signal a turn if they do not believe any road users are present to see the signal. For each of the factors associated with a driver’s odds of using a turn signal, an explanation might be considered that suggests that when other road users are likely present, signaling may be more likely. In a left-turn maneuver, there are more potential conflicts from vehicles approaching from the left, right, and opposing direction than on right turns where a conflicting vehicle approaches from the left. Perhaps because drivers perceive a greater risk in left-turn maneuvers, they are consequently 1.38 times more likely to signal a left turn. This result is consistent with Faw's (2013) observation that more signaling occurred in left turns compared to right. Converting Faw’s percentages to odds ratios, he observed that left turn signal use was about 1.45 times as likely as right turn signaling. This ratio falls within the confidence interval of the odds ratio observed in the present study. Traffic volume is likely to be greater on major and minor surface streets than on local roads. Such roads are also more likely to be signalized, contain multiple lanes of traffic, and have higher posted speed limits than local roads to accommodate greater traffic flow. While these roads certainly do not always have high traffic volume, they are likely to contain more traffic than local roads. One consequence might be the observation that signaling is five times more likely on major and minor surface roads than on local roads. This occurs despite the likely greater presence of designated turn lanes and protected signalized turns that have been associated with diminished turn signal use (Faw, 2013). This result also seems to contradict Faw’s observation that less turn signal use occurs in high-volume traffic, although, Faw’s traffic volume observations are not likely equivalent to the volume differences associated with
Table 4 Model fit statistics from three models. Model type
2 log likelihood AIC
4445.5 Intercept-only model All predictors 3688.6 Significant predictors only 3669.7
Pearson chi-square/DF
4447.5 1.00 3716.3 0.82 3689.8 0.78
Table 5 Odds ratios of turn signal use based on turn direction, road type, presence of a forward vehicle, and whether a stop was required before a turn. Comparison 1
Comparison 2
Direction Left
Right
Road type Major surface Major surface Minor surface
Minor surface Local Local
Odds ratio
Lower CL
Upper CL
0.003
1.38
1.12
1.69
1.0
Contents lists available at ScienceDirect
Accident Analysis and Prevention journal homepage: www.elsevier.com/locate/aap
Characteristics of turn signal use at intersections in baseline naturalistic driving John M. Sullivan a, *, Shan Bao a , Roy Goudy b , Heather Konet b a b
University of Michigan Transportation Research Institute, 2901 Baxter Rd., Ann Arbor, MI 48109-2150, USA Nissan Technical Center North America, 39001 Sunrise Drive, Farmington Hills, MI 48331-3287, USA
A R T I C L E I N F O
A B S T R A C T
Article history: Received 25 March 2014 Received in revised form 2 October 2014 Accepted 6 October 2014 Available online xxx
The purpose of this study was to determine whether a driver’s use of turn signals is sufficiently reliable to forecast a vehicle’s future path around an intersection, when detailed information about the intersection is unavailable. Naturalistic observations of turn signal use among 108 drivers on surface streets were extracted from the baseline portion of a field operational test of a safety system. Left and right turns that resulted in heading changes of between 70 and 110 and turn radii between 18 and 90 m were selected from the dataset. The odds that a driver would signal a turn were modeled as a function of road type, turn direction, presence of a forward vehicle, whether the vehicle stopped before the turn, and driver age and gender. Overall, 25 percent of left turns and 29 percent of right turns were not signaled. Road type, turn direction, and presence of a forward vehicle were found to influence the odds that a turn is signaled, while gender and age of the driver did not. The results suggest that situational factors like road type and turn direction are more powerful predictors of whether a turn will be signaled than either age or gender. Signaling on major and minor surface roads was about 5 times more likely than on local roads and 1.5 times more likely when a forward vehicle was present, suggesting a possible effect of traffic volume. It was concluded that turn signal activation alone may be insufficiently reliable to forecast a driver’s path. ã 2014 Elsevier Ltd. All rights reserved.
Keywords: Turn signal use Surface streets Turn direction Forward vehicle
1. Introduction In this study, naturalistic data were examined to help determine what systematic characteristics can be observed in a driver’s turning maneuvers that might be used to inform the development of a connected-vehicle-based intersection collision avoidance (ICA) application. Such an application would warn the driver not to proceed along a projected course if there is a high likelihood that a collision with another vehicle might occur. For such a system to do this well, it must be able to anticipate what the driver intends to do as the intersection is approached as well as determine what the other vehicles in the immediate vicinity are also likely to do. Prior attempts to address this type of road conflict have experimented with both vehicle-based radar to assess conflict (e.g., Pierowicz et al., 2000) as well as site-based radar approaches (Misener et al., 2010; Nowakowski, 2006; Penney, 1999). With the new capabilities enabled by connected vehicle technologies (also known as
* Corresponding author. Tel.: +1 734 764 8560; fax: +1 734 764 1221. E-mail addresses: [email protected] (J.M. Sullivan), [email protected] (S. Bao), [email protected] (R. Goudy), [email protected] (H. Konet). http://dx.doi.org/10.1016/j.aap.2014.10.005 0001-4575/ ã 2014 Elsevier Ltd. All rights reserved.
dedicated short range communications – DSRC), equipped vehicles within an approximately 300-m radius of each other will be able to transmit to and receive detailed information from other nearby vehicles. This form of wireless communication has been called vehicle-to-vehicle (V2V) communication to distinguish it from other kinds of connected-vehicle communications involving interactions between the vehicle and the local infrastructure (called V2I). Advantages of a V2V communications capability were noted more than 10 years ago by Miller and Huang, 2002 who proposed a peer-to-peer data sharing algorithm to be used to address intersection collision crashes with possible extensions to address both frontal and rear-end collisions. In their framework, the peer vehicles primarily shared position data derived from global positioning systems (GPS). Route contention was determined by calculating trajectories from the GPS data, and warnings were issued if the time-to-collision (TTC) fell to a value near to the time a driver was projected to require in order to avoid the collision. This time was called the time-to-avoidance (TTA). If the driver took any mitigating action (e.g., braking) before TTC became too low, the warning was withheld. Unlike an exclusively GPS-based system, V2V transmissions contain additional information that describes driver actions in nearby vehicles. Besides latitude, longitude, heading, and speed
2
J.M. Sullivan et al. / Accident Analysis and Prevention 74 (2014) 1–7
(available from on-board GPS), the basic safety message may also contain information about steering angle, braking, and turn signal use (see Society of Automotive Engineers, 2010 for message details). Thus, a collision avoidance system may be able to better predict a driver’s intention to stop, turn, or proceed through an intersection based on some of this additional information. A warning system’s reliance on this information will likely depend on how well it predicts what a driver will do in the immediate future. For example, recent work on turn signal use reports systematic variation in drivers’ use of turn signals. In two observational studies, Faw (2013) observed the turning maneuvers of drivers of passenger vehicles tabulating whether there was no signal, a late signal, or a good signal. In the first study, 3149 vehicles were observed executing both left and right turns at 47 different intersections; in the second study, selected based on low rates of turn signal use, 2455 right turns were observed at two intersections. Signaling was performed less often when turning right versus turning left, and less often when a dedicated turn lane was present. Turns at intersections were reported to be signaled about 76 percent of the time. This suggests that a turn is likely to be executed when a driver is signaling, but a turn may also be made 24 percent of the time when the driver is not signaling. Presence of traffic was also found to influence use of turn signals: in one study (Lebbon et al., 2007), 63 percent of drivers signaled turns in the presence of oncoming traffic compared to 44 percent when no traffic was present. On the other hand, Faw (2013) reports that turn signal use is lowest in heavy traffic (79 percent signaling), compared to light (89 percent) and moderate traffic (92 percent), suggesting less signaling in the presence of other traffic. The discrepancy might be a consequence of the small number of intersections sampled – Lebbon et al. (2007) sampled drivers at two intersections; Faw (2013) sampled drivers at 22 different intersections. Intersection characteristics and road configuration have also been noted to influence driver use of turn signals. For example, Faw (2013) reports 87 percent signaling when a dedicated turn lane is present, compared to 90 percent when no lane is present. The purpose of the present analysis is to examine naturalistic vehicle turning maneuvers for systematic regularities that could help indicate a driver's intention to turn as the driver approaches an intersection. This study examines turn signal use and its association with driver demographics (age and gender), and other factors derived from the naturalistic driving dataset including road type, turn direction, and the presence of lead vehicles on the roadway. Some of these factors could also be considered reasonable surrogates for traffic density – for example, it might be expected that arterial roads have a higher traffic density than local roads; likewise, the presence of a lead vehicle may be indicative higher traffic density than no lead vehicle. We also note that the naturalistic data differs in significant ways from the observational studies: driver age and gender are known; the drivers involved are all driving the same make and model vehicle (a 2006/2007 Honda Accord); and the set of 3830 turns in the sample were drawn from approximately 2732 unique locations. 2. Methods 2.1. Naturalistic data sample
(older), resulting in 18 drivers in each cell. Drivers were given test vehicles to drive unsupervised for a 12-day baseline period in which safety systems were inactive, followed by a 28-day period in which the safety systems were activated. The sample used for these analyses were derived from the baseline period of driving, when the safety systems were inactive. Thus, the baseline represents driving a vehicle equipped with no special warning systems that might artificially raise a driver’s awareness of the need to signal turns. 2.2. Turn selection criteria 2.2.1. Road type Because the targeted countermeasure scenario is intersection collision avoidance, selection of turn maneuvers was restricted to surface streets. This eliminated ramps, limited access roadways, and unknown road types which are largely parking lot maneuvers. Three road types were included in the analysis: major surface, minor surface, and local roads. Road types were derived from a commercial mapping database’s (NavTeq/HERE) five function classes and whether the roadway is identified as a limited access roadway or ramp. Major surface streets included roads supporting moderate-speed travel within cities and travel between cities (function class 2 and 3); minor surface streets included moderatespeed travel between neighborhoods (function class 4); and local roads were defined as supporting lower-speed travel between neighborhoods (function class 5). 2.2.2. Turn characteristics Turn maneuvers were selected that involved a heading change of between 70 and 110 . Turns were initially detected by examining vehicle yaw rates that exceeded 8 s 1. The beginning and end of the turn was determined by backtracking and forward-tracking in the yaw-rate time series until the absolute yaw rate dropped to about 0.5 s 1. This produced an approximate start and end of a turn (see the example turn shown in Fig. 1). Change in heading was determined by calculating the difference in heading at the start of the turn and the end of the turn. A negative difference was used to indicate a left turn; a positive difference indicated a right turn. After direct video inspection, several turns in the initial sample were found to occur in parking-lots or parking structures and were unrelated to intersection traversal; this occurred because the proximity of the parking area to a nearby road resulted in the
-83.4432 -83.4433 -83.4434 -83.4435 -83.4436 -83.4437 -83.4438 -83.4439
Turning data were developed from the driving data obtained in the Integrated Vehicle Based Safety System naturalistic driving study (Sayer et al., 2011). In this study, 108 licensed drivers were recruited to drive passenger vehicles equipped with a combination of safety systems. The sample of drivers was equally divided between male and female drivers and stratified into three age groups: 20–30 years (younger), 40–50 (middle-aged), and 60–70
-83.444 42.5449 42.545 42.545142.545242.545342.545442.545542.545642.545742.5458
Fig. 1. GPS trace of path of a vehicle executing an approximately 90 turn; the red circles identify the part of the turn examined in the analyses. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)
J.M. Sullivan et al. / Accident Analysis and Prevention 74 (2014) 1–7
3
Fig. 2. Frequency distribution of turn radii among the selected curves. The red vertical line is the mean of each distribution. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)
vehicle being erroneously located on one of the acceptable road classes. In addition, some 70–110 changes in direction also occurred on uninterrupted roadways and did not involve intersection maneuvers at all. This usually occurred along rural roadways. To address these issues, turns were further filtered by determining whether the turn occurred within 5 m of an identified roadway node, using State of Michigan roadway geometric data derived from the Center for Geographic Information (http://www. michigan.gov/cgi). One consequence of this filtering was that driving that occurred outside the State of Michigan was excluded from the sample. Turns were also filtered based on average turn radius, calculated from the total distance traveled over the turn and
the total heading change. Turns with a radius less than 18 m were most often observed in parking areas and residential driveways; turns with a radius of more than 90 m were most often observed on continuous rural roadways and not related to intersections. These criteria were initially established based on the Delaware State Department of Transportation design criteria for the inside radius of intersections having a design speed of between 24 and 56 km/h (15–35 mph) (Delaware Department of Transportation, 2006). Direct video inspection was used to confirm this selection strategy. Following application of these filtering criteria, a pool of 3830 turns remained in the data sample, of which 2732 occurred at unique locations.
Fig. 3. Length of turn by direction and road type. The red vertical line identifies the mean of each distribution. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)
4
J.M. Sullivan et al. / Accident Analysis and Prevention 74 (2014) 1–7
Fig. 4. Duration of turn by direction and road type. The red vertical line identifies the mean of each distribution. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)
3. Results 3.1. Characteristics of sample turn paths The sample of turns was characterized with respect to distribution of continuous variables related to turn execution – turn radius, turn length, turn duration, heading change – over categorical factors that included turn direction, road type, driver age group, and gender. Each of the continuous variables was modeled as a dependent measure using the categorical factors as predictors in a mixed model analysis that included driver as a random factor. The statistics reported here describe the Bonferroni-adjusted pairwise comparisons between factors in which systematic differences were observed for each dependent measure. In general, the turn radii of local roads were systematically shorter by 2–3 m than either major or minor surface streets (t = 3.3, p < 0.01, and t = 5.0, p < 0.01, respectively; see Fig. 2), and the turn radii of right turns was shorter than left turns by about 1.6 m (t = 4.2, p < 0.01). These radii were based on the actual turn paths that drivers followed while executing turns and do not reflect the geometric design of the roadway. No effect of driver age or gender was observed. Consistent with the turn radii data, the distribution of turn length was also systematically shorter on local roadways than major and minor surface streets and shorter for right turns than left turns (shown in Fig. 3). The turn length along local roads was observed to be an average of 3–5 m shorter than major or minor surface roads (t = 3.1, p < 0.001, p < 0.001, t = 5.4, p < 0.001, respectively). The average length of right turns was about 3 m shorter than left turns (t = 4.8, p < 0.001). No effects of driver age or gender were observed. The duration of turn execution averaged between 7 and 8 s, taking about 0.4–0.5 s longer from local roads than from major or minor surface streets (t = 5.5, p < 0.01; t = 5.6, p < 0.01, respectively). No systematic differences were observed in turn duration between left and right turns (see Fig. 4). Younger drivers traversed turns in about 0.45 s less time than middle aged and older drivers (t = 4.0, p < 0.01; t = 4.1, p < 0.01).
The distribution of heading changes after selection of turns conformed to the approximate 90 left and right heading changes expected after application of the turn selection strategy. No systematic differences were observed in the sample heading changes were across the factors (Fig. 5). 3.2. Sample driver distribution While care was taken to ensure that there was a relatively balanced distribution of male and female drivers of each age group that participated in the IVBSS study, the study did not regulate drivers’ exposure to turns. Drivers who logged more hours driving were likely to execute turn maneuvers more often than drivers who did not. The proportional makeup of the driver sample is shown in Table 1. The sample of turns has a greater proportion of young drivers than middle aged and older drivers. Within the young and middle aged categories, proportionally more turns were executed by male drivers than female drivers; the gender split between the older drivers is approximately even. The distribution of gender and age is not independent (x2 = 17.8, p < 0.001). The generalized linear mixed model analyses presented later includes driver as a random factor to ensure that results are not biased by disproportionate numbers of observations by driver. The distribution of driver age across turn direction was also examined in light of prior work that suggests older drivers may self-limit their driving (Braitman and Williams, 2011), perhaps to avoid complicated roadway situations. The sample was examined for evidence of any particular age-related selectivity in turning maneuvers. As can be seen in Table 2, no evidence was found to suggest that turn distributions among older drivers differed from younger and middle aged drivers (x2 = 0.11, p = 0.94). 3.3. Driver’s use of turn signals For each turn in the dataset, turn signal use was identified for all cases in which the turn signal was actuated at some point within the turn execution. For most turns, activation occurred before the turn began, although, drivers occasionally began signaling after the start of the turn. For the present analysis, both kinds of signal
J.M. Sullivan et al. / Accident Analysis and Prevention 74 (2014) 1–7
5
Fig. 5. Heading change distribution by direction and road type. The red vertical line identifies the mean of the heading change distribution. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)
vehicle. This eliminated most stationary targets, vehicles traveling toward the subject vehicle, and vehicles moving away from the subject vehicle at more than twice the subject vehicle’s current speed. The intention of this selection strategy was to identify situations in which the driver’s approach to an intersection might be constrained by the presence of other vehicles in proximity to the subject vehicle. Secondarily, the indicator identifies situations in which forward traffic is present as the intersection is approached. Vehicles that stopped before the turn were identified as those whose speed was less than 1 m s 1 any time in the 100-m distance before the start of the turn. Vehicles commonly stop in response to traffic signals or other traffic impediments like cross-intersection traffic.
actuation are considered as signaling. The breakdown of turn signal use in the study sample is shown in Table 3. Overall, unsignaled turns occurred in about 25 percent of the left turns, and 29 percent of the right turns. Signaling errors, in which a turn is made in the direction opposite to the direction of the turn executed, were quite rare – about 0.1 percent. One of these cases resulted from a left turn off a right-curving road that was misidentified by the analysis algorithm. The other two cases involved one driver who appeared to abruptly change their mind about the turn direction (e.g., their gaze was initially toward the signaled direction but abruptly changed), and another who signaled to exit a left-turn lane, abandoned the attempt, and turned left. Generalized linear mixed models were used for this part of the analysis and account for repeated observations of the same driver. The between-subject variables were age (young, middle, old) and gender (male, female). The four within-subject variables were roadway type (local, major surface, minor surface), direction of the turn (left, right), presence of a forward vehicle moving in the same direction as the subject vehicle within a distance of 100 m (yes, no), and whether the turn was preceded by a stop (yes, no). In the mixed model, driver and interactions between the driver and within-subject effects were treated as random effects. This accounts for within-subject correlated observations and effectively allows comparison of a driver to him/herself. The dependent variable in this analysis was whether a driver signaled the turn. Presence of a forward vehicle was determined by scanning the vehicle’s approach to the start of the turn (beginning 100 m back) for the presence of any targets in the forward radar within 100 m and traveling at a relative absolute speed less than the subject
3.3.1. Analysis results Three models were constructed and their fit statistics compared. An intercept-only model was used to establish a base model with none of the predictors present. A second model was constructed that included all of the predictors. And finally, a third model was evaluated that included only predictors in which a significant association with the dependent measure was observed. The fit statistics of each model are shown in Table 4 for comparison. The results suggest that the fit of the model that includes the significant predictors is best, and is the one selected for discussion. Main effects were observed for turn direction, road type, and forward vehicle presence. No effects were observed for driver age group or gender. A marginally significant effect was also observed in situations in which a vehicle stopped before initiating the turn. Odds ratio estimates are shown in Table 5. Left turns were about 1.38 times as likely to be signaled as were right turns. There was
Table 1 Distribution of sample curves among drivers in sample.
Table 2 Distribution of sample turns by driver age and turn direction.
Age group
Age group
Gender
Young
Middle
Older
Total
Turn direction
Young
Middle
Older
Total
Male Female Total
906 637 1543
646 506 1152
574 561 1135
2126 1704 3830
Left Right Total
746 797 1543
557 595 1152
542 593 1135
1845 1985 3830
6
J.M. Sullivan et al. / Accident Analysis and Prevention 74 (2014) 1–7
Table 3 Turn signal use by turn direction. Turn direction
Counts
Signal status
Counts
Percent
Left
1845
Unsignaled Signaled left Signaled right Unsignaled Signaled left Signaled right
453 1390 2 570 1 1414
24.6 75.3 0.1 28.7 0.1 71.2
Right
Total turns
1985
3830
little difference in the odds of signaling between major and minor surface streets, however each differed from the odds of signaling on local streets by more than 4 times. When a forward vehicle was present, signaling was about 1.5 times as likely as it was when no forward vehicle was present. Finally, signaling was also about 1.3 times as likely if a vehicle had stopped in the 100 m preceding the turn. 4. Discussion One interpretation of the observed results is that they may be related to the likelihood that other vehicles are present on the roadway to motivate the driver to communicate the intended maneuver. On sparsely populated roadways, drivers may be less inclined to signal a turn if they do not believe any road users are present to see the signal. For each of the factors associated with a driver’s odds of using a turn signal, an explanation might be considered that suggests that when other road users are likely present, signaling may be more likely. In a left-turn maneuver, there are more potential conflicts from vehicles approaching from the left, right, and opposing direction than on right turns where a conflicting vehicle approaches from the left. Perhaps because drivers perceive a greater risk in left-turn maneuvers, they are consequently 1.38 times more likely to signal a left turn. This result is consistent with Faw's (2013) observation that more signaling occurred in left turns compared to right. Converting Faw’s percentages to odds ratios, he observed that left turn signal use was about 1.45 times as likely as right turn signaling. This ratio falls within the confidence interval of the odds ratio observed in the present study. Traffic volume is likely to be greater on major and minor surface streets than on local roads. Such roads are also more likely to be signalized, contain multiple lanes of traffic, and have higher posted speed limits than local roads to accommodate greater traffic flow. While these roads certainly do not always have high traffic volume, they are likely to contain more traffic than local roads. One consequence might be the observation that signaling is five times more likely on major and minor surface roads than on local roads. This occurs despite the likely greater presence of designated turn lanes and protected signalized turns that have been associated with diminished turn signal use (Faw, 2013). This result also seems to contradict Faw’s observation that less turn signal use occurs in high-volume traffic, although, Faw’s traffic volume observations are not likely equivalent to the volume differences associated with
Table 4 Model fit statistics from three models. Model type
2 log likelihood AIC
4445.5 Intercept-only model All predictors 3688.6 Significant predictors only 3669.7
Pearson chi-square/DF
4447.5 1.00 3716.3 0.82 3689.8 0.78
Table 5 Odds ratios of turn signal use based on turn direction, road type, presence of a forward vehicle, and whether a stop was required before a turn. Comparison 1
Comparison 2
Direction Left
Right
Road type Major surface Major surface Minor surface
Minor surface Local Local
Odds ratio
Lower CL
Upper CL
0.003
1.38
1.12
1.69
1.0
