Improving Bounds on Occluded Vehicle States for Use in Safe Motion Planning

Sensor occlusions pose a large risk to autonomous vehicles due to the unknown nature of what lies within. Previous research has introduced methods for predicting the future occupancy of hidden vehicles. Potential hidden vehicles are represented as continuous intervals on position, angle, and speed. These intervals are defined based on sensor occlusions, but over-represent true possible hidden vehicle state. The resulting occupancy predictions are restrictive to motion planning, and cause unwanted over-cautious behavior. Therefore, we introduce our method for reducing the bounds on the state interval used in prediction. We detail how tracking occlusion motion can be used to restrict the position interval, and observation over time can be used to find an upper limit on the exit velocity. The smaller initial intervals yield much smaller predicted occupancies, especially over long prediction horizons. Our method is evaluated in simulation alongside the previous method, using a basic provably-safe planner. Over three simulated scenarios, we show that our approach is able to navigate intersections up to 3 seconds faster, and avoid unnecessary braking maneuvers.

I. Introduction

Maintaining public trust in autonomous vehicles is critical to the adoption and ultimate success of the industry [1]. Autonomous Vehicles with quantifiable prediction accuracy and provably safe planning have the potential to drastically improve safety on the road [2]. Given the potential benefits, researchers say it is in the public interest to deploy soon and expedite adoption [3]. Research and investment in autonomous vehicles has grown considerably in the last 10 years. Technical advancements in the last 5 years have led to more than three times as many performing on-road testing, over longer distances, and with fewer disengagements per mile. Additionally, the autonomous vehicle industry is projected to grow by an order of magnitude to a value of over $500 billion USD by 2026 [4]. There is clearly potential for widespread adoption, however, despite projections and technological advancement, there are still concerns about safety.