Towards Non-Flat Road Perception: Simulation and Analysis

Semantic segmentation of point cloud is an important part of environmental perception for autonomous vehicles. However, a large number of existing open-source point cloud datasets lack data for non-flat road environments such as rural areas. These road scenes have complex characteristics, and the models trained with traditional urban datasets cannot perform well in such environments for semantic segmentation. In this paper, we utilize two methods, the Gaussian function and the imitation generation method, to simulate non-flat road environments. Based on training using the simulated dataset, we design a new point cloud-based semantic segmentation method for non-flat road environments. Finally, through multiple experiments, we demonstrate that the proposed semantic segmentation model achieves good results in non-flat road environments and exhibits generalizability when faced with different road conditions.