An occlusion- and interaction-aware safe control strategy for autonomous vehicles

Abstract

We consider the problem of safe autonomous driving in the presence of occlusions. Dealing with latent risks arising from occlusions is challenging because there does not exist direct mapping from sensor input to visible threats; attempts to ensure safety for all worst-case latent threats can be infeasible or overly conservative, and accounting for a multitude of latent risks for sufficient future horizon may require prohibitive computation in real-time. To address these issues, in this paper, we propose to use a probability-based predictive controller to make safe decisions for autonomous vehicles. We prove that the proposed safety controller can generate vehicle control profiles that yield the desired safety probability. Numerical and onboard experiments on a visual occluded pedestrian crossing scenario verifies the efficacy of the proposed method in real-time. The merits of the proposed control strategy include being able to guarantee long-term safety under occlusions without being over-conservative, handling latent risks caused by on-road interactions in real-time, and ease of design with transparency to the exposed risks.