Efficient Uncertainty-Aware Collision Avoidance for Autonomous Driving Using Convolutions

Motion planning directly in the spatiotemporal dimension can generate trajectories of higher quality compared to decoupled methods for autonomous driving. However, it requires a greater amount of computational resources. This paper proposes an efficient motion planning method based on convolution in the spatiotemporal dimension, which takes into account the uncertainty of localization and obstacle intention. Firstly, a three-dimensional probability occupancy grid map with uncertainty is constructed based on prediction results. Secondly, convolution kernels are generated considering the contour, heading angle and localization uncertainty of the ego vehicle. Thirdly, single-channel multi-output convolutions are performed between the probability occupancy grid map and the kernels to generate the four-dimensional feature map. Finally, a collision avoidance algorithm based on the feature map is proposed to obtain the optimal trajectory, which uses the hybrid A* algorithm. The chance constraint and the vehicle kinematics are taken into account in the motion planning. In simulation experiments, the safety performance, computational efficiency and rationality of the motion planning are compared and analyzed, and the proposed method performs superiorly. In addition, real-world experiments verify the feasibility of the proposed method.