Modelling pedestrian behavior at road intersections using reinforcement learning frameworks

Abstract: Understanding the behavior and intent of pedestrians and other vulnerable road users is vital for ensuring safety at road intersections. Despite this, there is a notable gap in the modeling and prediction of the inherently erratic nature of pedestrian movements and interactions. Addressing this challenge, our study employs non-identifiable LiDAR systems to gather comprehensive data on pedestrian and vulnerable road user behaviors across multiple intersections at varying times of the day. This approach allows us to track precise positional information and develop reinforcement learning agents capable of replicating observed movement patterns. By simulating these patterns, we aim to create predictive models that provide valuable insights into pedestrian behavior when interacting with vehicles. We go beyond simple modeling by evaluating the synthesized movement patterns against actual vehicle tracking data to forecast the intentions of pedestrians and vulnerable road users in a controlled digital environment. This approach enables us to test the accuracy and adaptability of the reinforcement learning agents under diverse synthetic scenarios, ensuring that the models accurately represent real-world behaviors. Our findings indicate that these agents are effective in mimicking pedestrian dynamics, contributing critical insights into potential conflict points at intersections. The implications of our research are significant. We anticipate that our results will serve as a foundational tool for enhancing intersection safety protocols and inform innovative intersection design strategies. These insights could be pivotal in progressing toward a vision of zero incidents involving pedestrians and vulnerable road users. By refining our understanding of pedestrian behavior and integrating it with intelligent intersection management, we hope to contribute to safer urban environments and proactive road safety measures.