GPNet: Point Cloud Semantic Segmentation Network for Off-Road Scenes

Semantic segmentation of point clouds plays a critical role in autonomous driving as it enables the processing of lidar-obtained data to achieve environmental perception. While significant progress has been made in the semantic segmentation of point clouds in urban traffic environments, research on off-road scenes has been limited due to a lack of datasets. In this paper, we present a semantic segmentation model for point clouds, tailored to address the unique challenges of off-road environments. Given the characteristics of off-road scenes, such as indistinct boundaries, uneven terrain, irregular features, and relatively few occlusions, our model utilizes a projection method to segment point cloud data into Bird's Eye View (BEV) and extracts adjacent features through adaptive graph convolution to compensate for lost information during projection. Our model achieves high performance on Rellis-3D and SemanticKITTI datasets.