Kinematic cues in driver-pedestrian communication to support safe road crossing.

OBJECTIVE: Right-of-way negotiation between drivers and pedestrians often relies on explicit (e.g., waving) and implicit (e.g., kinematic) cues that signal intent. Since effective driver-pedestrian communication is important for reducing safety-relevant conflicts, this study uses information theory to identify vehicle kinematic behaviors that provide the greatest information gain and serve as cues for pedestrians to cross safely. DATA SOURCES: A driver-pedestrian dataset with 348 interactions was extracted from a large naturalistic driving data collection effort. It includes 325 instances of a pedestrian crossing the vehicle's path and 23 instances in which the vehicle did not yield to a pedestrian. Kinematic data were collected from the vehicle's CAN. Pedestrian behaviors, driver cues, and contextual information were manually annotated from a forward-facing video. METHODS: We used kernel density estimation to quantify the probabilities of vehicle acceleration, speed, and standard deviation of speed, for a given vehicle position and pedestrian behavior. Mutual information was then calculated between the estimated distributions given a pedestrian behavior (crossing/not crossing; walking/pausing) across intersection types (protected, e.g., stop signs; designated, e.g., crosswalks; and undesignated, e.g., jaywalking). RESULTS: The patterns mutual information conveyed by vehicle kinematics differed across measures (acceleration, speed, and standard deviation of speed) reaching peak values (in bits of information) at different distances from the pedestrian path. The mutual information conveyed by vehicle acceleration and pedestrian crossing behaviors peaked the farthest from the pedestrian path in the designated crossings, about 18 m away from the pedestrian path, with a difference in median deceleration of 1.01 m/s2 (p < 0.001) between pedestrian pausing and walking epochs. For protected crossings, the peak in mutual information occurred closer (10 m) to the pedestrian path, where median vehicle deceleration was significantly lower (0.55 m/s2; p < 0.05) in pausing epochs compared to walking epochs. For undesignated crossings, the peak in mutual information was the closest to the pedestrian crossing path, around 5 m, and was associated with a stronger deceleration behavior in pedestrian crossing epochs (-0.33 m/s2; p < 0.1). Vehicle speed demonstrated a similar sensitivity to distance from the pedestrian path across intersection types. Lastly, looking at the outcome of pedestrian behavior (i.e., crossing/not crossing), we find that the mutual information conveyed by acceleration, speed, and standard deviation of speed, peaked when the vehicle was at 30 m (stronger braking -0.37 m/s2; p < 0.1) and 10 m away, with greater acceleration (0.81 m/s2; p < 0.001) and faster speeds (2.41 m/s; p < 0.001) in pedestrian crossing epochs. SIGNIFICANCE OF RESULTS: This study examined driver-pedestrian information exchange using vehicle kinematic behavioral cues. We find that the differences in mutual information are shaped by multiple factors including the intersection type. In general, there was less mutual information gain in protected crossings which may be explained by unambiguous right-of-way rules guiding driver and pedestrian behavior, reducing the need for negotiation. Driver-pedestrian interactions in designated crossings seem to take place over a larger distance range compared to undesignated or protected crossings. These findings may support the design of automated driving and pedestrian safety systems that are able to consider the type, strength, and timing of kinematic cues to optimize driver-pedestrian negotiation. Eventually, such systems may enhance safe, efficient, and social interactions with pedestrians.

Interdependence of driver and pedestrian behavior in naturalistic roadway negotiations.

OBJECTIVE: This paper characterizes the actions of pedestrian-driver dyads by examining their interdependence across intersection types (e.g., zebra crossings, stop signs). Additionally, the analysis of interdependence captures other external factors, such as other vehicles or pedestrians, that may influence the interaction. METHODS: A 228 epoch vehicle-pedestrian interaction dataset was extracted from a large naturalistic driving data collection effort, which included vehicle, pedestrian, and contextual information (e.g., intersection type, jaywalking, vehicle maneuver, and lead vehicle presence). An expanded Actor-Partner Interdependence Model (APIM) was used to analyze driver-pedestrian dyads using driver and pedestrian standard deviations of velocity as the independent variables and wait times as dependent variables. APIM structural equation models were augmented to include driver effects (i.e., lead vehicle and maneuver type) and pedestrian effects (i.e., lead pedestrian, crossing group size, crossing direction). RESULTS: The level of protection afforded by an intersection had an effect on the extent of driver-pedestrian dyadic behavior. Interactions in undesignated crossings (i.e., jaywalking) were associated with interdependent behavior whereas interactions in designated crossings (i.e., crosswalks and parking lots) showed a partner effect on the driver's wait time but no significant corresponding partner effect on the pedestrian. Finally, protected intersection interactions (i.e., traffic lights and stop signs) demonstrated no significant partner effects. CONCLUSIONS: The difference in behavior patterns associated with the intersection type and level of protection shows that context can mediate the level of negotiation required between drivers and pedestrians. These findings inform how context and driver-pedestrian interactions should be incorporated in future modeling efforts which may, ultimately, support design of automated systems that are able to interact more safely, efficiently, and socially.