Multi-agent trajectory prediction at unsignalized intersections: An improved generative adversarial network accounting for collision avoidance behaviors

Abstract

Accurate trajectory prediction for multiple agents (i.e., vehicles, bicyclists, and pedestrians) is the premise of launching proactive interventions, which can serve as an effective way to improve traffic safety at unsignalized intersections. The distinctive characteristic of unsignalized intersections lies in their disorderly traffic organization, prompting traffic agents to be extra vigilant towards other agents to prevent collisions. As such, the primary focus of multi-agent trajectory prediction lies in acquiring a deep understanding of their interactive behavior patterns when encountering potential collisions. To achieve this, this study proposes an improved generative adversarial network (GAN) that can properly model collision avoidance behaviors of multiple agents when predicting their trajectories. Specifically, attention pooling modules are employed to capture interactions among multiple agents. A graph convolution network (GCN) based collision extraction module is applied to identify potential collisions and model the collision avoidance behaviors of traffic agents. Experimental results on InD dataset demonstrate that the proposed framework attained a more accurate and reliable performance for multi-agent trajectory prediction. In different interactive scenarios, such as when vehicles yield or don’t yield, the results illustrated via the Distance-velocity (DV) diagram display a significant level of robustness. Furthermore, the conflict points, count of dangerous interactions, and Post-Encroachment Time, as computed from these predicted trajectories, also align well with the ground truth. This indicates that the proposed framework effectively captures the pattern of collision avoidance behaviors of multiple agents, which has potential to serve as an effective way to enhance traffic safety at unsignalized intersections.