Contingency-Aware Intersection System for Autonomous and Human-Driven Vehicles with Bounded Risk

Abstract

Traffic intersections are natural bottlenecks in transportation networks where traffic lights have traditionally been used for vehicle coordination. With the advent of communication networks and Autonomous Vehicle (AV) technologies, new opportunities arise for more efficient automated schemes. However, with existing automated approaches, a key challenge lies in detecting and reasoning about uncertainty in the operating environment. Uncertainty arises primarily from AV trajectory tracking error and human-driven vehicle behavior. In this paper, we propose a risk-aware intelligent intersection system for AVs along with human-driven vehicles. We formulate the problem as a receding-horizon Chance-Constrained Partially Observable Markov Decision Process (CC-POMDP). We propose two fast risk estimation methods for detecting vehicle collisions. The first provides a theoretical upper bound on risk, whereas the second provides an empirical upper bound and runs faster, hence more suitable for real-time planning. We examine our approach under two scenarios: (1) a fully autonomous intersection with AVs only, and (2) a hybrid of signalized intersection for human-driven vehicles along with an intelligent scheme for AVs. We show via simulation that the system improves the intersection's efficiency and generates policies that operate within a risk threshold.