Predictive Trajectory Planning for On-Road Autonomous Vehicles Based on a Spatiotemporal Risk Field

Abstract

Trajectory planning plays a crucial role in autonomous driving. It generates feasible geometric paths within a certain time horizon to guide autonomous vehicles. However, typical planning methods give little consideration to the motion uncertainty of traffic objects. Therefore, this article proposes a predictive trajectory planning method based on a spatiotemporal risk potential field. And the key notion of this method is to separate the coupled executions, i.e., trajectory generation and risk assessment, into two parallel and independent modules in the curvilinear coordinate. First, the current spatial risks are quantified by artificial potential-like functions. Then, we measure the probabilities of surrounding vehicles driving toward each lane using an interacting multiple model structure. The temporal evolution of the spatial field is computed by incorporating both the probabilities and predicted trajectories. Meanwhile, the generation module focuses on providing sufficient trajectory candidates for the ego vehicle. To allocate computing resources reasonably, feasibility probabilities are calculated with a relational grid, and then candidates are biased sampled in lateral and longitudinal directions. Finally, taking into account driving safety, comfort, and efficiency, the planner selects an optimal trajectory by minimizing a well-designed cost function. In this way, the feasibility of planned trajectories can be greatly improved with limited time consumption. The performance of this proposed method is evaluated in various scenarios using a high-fidelity Unity3D environment.