LiDAR-based Ground Segmentation Using Stixel Features

For autonomous vehicle, ground segmentation is one of the most basic and essential capabilities. In this paper, to further improve the performance of LiDAR-based ground segmentation approaches, other than the commonly used `oc-cupancy' information of the point clouds, we try to utilize the `free' information naturally contained in any point clouds. We choose to use the stixel as the processing primitives. By utilizing the intrinsic relationship between the 3D voxel grid and the 2D range image, we can efficiently calculate the stixel features that contain both the occupancy information and the free information. Using the stixel features as inputs, we design a deep learning network for ground segmentation. The designed network consists of a feature extraction net and an ERF (Efficient Residual Factorized) semantic segmentation net. Experiments on SemanticKITTI dataset show that our proposed approach outperforms state-of-the-art approaches.