Road Grid Segmentation in Desert with Geometry and Appearance Features
Yinfeng Zhao, Tao Wu, Meiping Shi, Xueyan Liu, Xijun Zhao, Yanchen Hu · 2023
Presently, the detection of drivable areas is one of the most important tasks for autonomous driving of UGV (unmanned ground vehicle). In off-road scenes, the accessible areas have complex appearance. There is no obvious boundary between obstacles and roads. Besides, each single frame of Lidar scans is greatly affected by the slope of ground. Therefore, this article proposes a road segmentation method based on LiDAR. Firstly, through multi-frame accumulation and Gaussian filtering, a local ground height map is obtained. Then we extract slope information based on the height map. The current scanned point cloud appearance features and slope features are fused to achieve ground grid segmentation in a bird's eye view. In training, this article uses dense labels obtained from multi-scans to supervise sparse scan scenes. Experiments has shown that the fusion of the two features improved the accuracy of the algorithm. This paper also collected 2500 frames of point cloud in Alax, and created an off-road dataset for the Gobi Desert. On this dataset, our method outperforms several comparison methods and can be implemented in real-time operation.