Autonomous Vehicles’ Turning Motion Planning for Conflict Areas at Mixed-Flow Intersections

Abstract

An essential and challenging task for autonomous vehicles (AVs) is turning at mixed-flow intersections when they interact with motorized, non-motorized and pedestrian traffic simultaneously. In these situations, sensor measurement noise and blind zones create additional complexity in an already problematic environment. In order to make motion planning feasible in a multi-interaction environment with detection uncertainty feasible, this article proposes a hierarchical framework that divides the highly-related driving process into a decision, planning and action layer. The decision layer first employs a logit model combined with Bayes' theorem to make a discrete choice about whether to turn or not. Then, the plan layer initializes a local trajectory with selected waypoints considering the location of interacting objects based on a Bézier curve. Finally, feedback is used to adjust the vehicle's decision and trajectory plan when collision risks increase due to the unexpected behavior of other objects. Additionally, in order to consider sensor data noise and blind zones of AVs, an Extended Kalman filter (EKF) was used to estimate the status of sensory targets. The performance of the proposed model was compared with drivers' performance for the same turning scenarios at two mixed-flow intersections. The results show that the simulation produced human-like flexibility when planning trajectories in a multi-interaction environment. Moreover, the travel time of AVs did not show a statistically significant difference when compared with manually driven vehicles (MVs). Instead, the AVs actually performed better in terms of safety than the MVs.