A convolution-based motion planning method for autonomous driving with localization uncertainty

This paper proposes a convolution-based motion planning method for autonomous driving, where the uncertainty of vehicle localization is modelled and incorporated. The well-known concept of "convolution" in artificial intelligence is integrated into the motion planning. Vehicles with possibly different heading angles are taken as kernels, and each kernel is assigned a grid occupancy probability based on the uncertainty of the localization. Multiple convolutions between the kernel and the environment are performed in advance to generate a 3D feature map, which significantly increases the computational efficiency of collision detection algorithms in motion planning. The obtained trajectories are verified to be more secure and reliable under highly uncertain localization conditions in comparison simulations. Copyright (C) 2023 The Authors.