Improved Occlusion Scenario Coverage with a POMDP-based Behavior Planner for Autonomous Urban Driving

Abstract

Safely driving through various occlusion scenarios in urban environments, such as bus stops or crosswalks, is challenging for autonomous vehicles (AVs). Improving the ability to handle more occlusion scenarios in urban environments is paramount when using AVs as shuttle buses. An AV could experience deadlock situations in very heavy occlusion scenarios with the worst-case assumption that potential occluded road users could suddenly emerge using maximal allowed velocity. In this study, we address this issue with a partially observable Markov decision process (POMDP)-based behavior planner to improve the occlusion scenario coverage. We extend a phantom vehicle concept to include pedestrians to represent potential road users in risky occlusion areas. The appearance probability of phantom objects along with their future movement is inferred using map information and road topology. Finally, context-aware phantom road users are incorporated within a POMDP formulation, which is solved online by constructing a Monte Carlo tree with reachable state analysis. Various evaluation results indicate that the ego vehicle shows comfortable driving behavior, aiming to avoid unnecessary braking and acceleration when driving through challenging occlusion scenarios in urban areas, including crosswalks, bus stops, and intersections. Moreover, it does not lead to deadlock situations in heavily occluded scenarios.