Graph Convolution Network for Road Detection with Lidar

Road detection is an important task in environment perception of intelligent driving. The road detection method is usually affected by many aspects, such as the occlusion of road vehicles and pedestrians, the change of illumination, etc. It is noting that lidar is insensitive to light and perceive the 3D environment information around the vehicle well during the day and night. So lidar road detection is suitable for intelligent driving scenarios. In this work, a Graph Convolution Network, named GCN RoadNet, is proposed for road detection using lidar point cloud, which discriminates on-road points from off-road points. Graph convolution network focuses more on local connections of point cloud points than global ones, therefore it achieves better accuracy than PointNet. Experiments on the KITTI dataset validate the robustness and effectiveness of the proposed method.