A LiDAR Semantic Segmentation Framework for the Cooperative Vehicle-Infrastructure System

LiDAR semantic segmentation plays an important role in 3D scene understanding for autonomous driving. However, the performance based on the LiDAR equipped on a vehicle may be limited due to small perception perspective, object occlusion, and sparsity of point clouds in the distance. To address these challenges, we propose a vehicle-infrastructure cooperative semantic segmentation (VICSS) framework to enhance the vehicleside perception capability. An infrastructure feature extraction (IFE) module is employed to extract features from the roadside LiDAR point cloud. A local feature extraction (LFE) module and a global feature extraction (GFE) module are applied to the vehicle point cloud to capture the local and global features, respectively. The features from both point clouds are fused through a feature aggregation (FA) module, which applies the cross-attention mechanism to learn beneficial information from the infrastructure features. Using a dataset generated by CARLA we show that the proposed VICSS framework achieves good performance in terms of semantic segmentation accuracy.