Predicting pedestrian road-crossing assertiveness for autonomous vehicle control

Abstract

Autonomous vehicles (AVs) must interact with other road users including pedestrians. Unlike passive environments, pedestrians are active agents having their own utilities and decisions, which must be inferred and predicted by AVs in order to control interactions with them and navigation around them. In particular, when a pedestrian wishes to cross the road in front of the vehicle at an unmarked crossing, the pedestrian and AV must compete for the space, which may be considered as a game-theoretic interaction in which one agent must yield to the other. To inform AV controllers in this setting, this study collects and analyses data from real-world human road crossings to determine what features of crossing behaviours are predictive about the level of assertiveness of pedestrians and of the eventual winner of the interactions. It presents the largest and most detailed data set of its kind known to us, and new methods to analyze and predict pedestrian-vehicle interactions based upon it. Pedestrian-vehicle interactions are decomposed into sequences of independent discrete events. We use probabilistic methodslogistic regression and decision tree regression-and sequence analysis to analyze sets and sub-sequences of actions used by both pedestrians and human drivers while crossing at an intersection, to find common patterns of behaviour and to predict the winner of each interaction. We report on the particular features found to be predictive and which can thus be integrated into gametheoretic AV controllers to inform real-time interactions.

Pedestrian Models for Autonomous Driving Part II: High-Level Models of Human Behavior

“…3) Feature selection methods: Bonnin et al [27] proposed a generic context-based model to predict pedestrians behavior according to features describing their local urban environment. To learn about interactions between autonomous vehicles and pedestrian interactions, in [37], Camara et al collected data from real-world pedestrian-vehicle interactions at an unsignalized intersection. The actions of pedestrians and vehicles were ordered into sequences of events comprising descriptive features and the study revealed the most predictive features in a crossing scenario such as the head direction, the position on the pavement, hand gestures etc.…”

“…This is consistent with recent research [166] that reported head orientation/gaze towards vehicles as the most prominent cues for predicting pedestrian intent. In addition, computational models have shown that head direction is a useful trait for pedestrian path prediction and state of situation awareness such as in [37] which argued that if a pedestrian looks at the vehicle, they are less likely to cross the road.…”

“…Social ground truth annotations are much rarer. [36] and [37] collected high quality human annotations of physical and social events during pedestrianvehicle interactions, including the presence and timings of the agents communicating with each other via eye contact, hand gestures, positions and speeds, and the final 'winners' of interactions which compete for road space during crossings.…”

Pedestrian Models for Autonomous Driving Part I: Low-Level Models, From Sensing to Tracking

“…The most basic method of sensing pedestrians is to use human perception, which is often used in offline studies, such as for conducting on-street surveys or annotate recordings of such surveys made with other sensors [29], [30]. Humans still have advantages over automated systems since they can use their full intelligence to subjectively annotate otherwise difficult events, such as the meanings of body language, emotions, and gestures.…”

Game Theory For Self-Driving Cars

“…We collected a large scale data from real-world human road crossings at the intersection near the University of Leeds, UK. Pedestrian-vehicle interactions were decomposed into sequences of independent discrete events [9]. We looked for common patterns of behaviour that can predict the winner of an interaction, which can thus be integrated into gametheoretic AV controllers to inform real-time interactions.…”