TCFNet: Transformer and CNN Fusion Model for LiDAR Point Cloud Semantic Segmentation

Abstract

Dynamic scene understanding based on LiDAR point clouds is one of the critical perception tasks for self-driving vehicles. Among these tasks, point cloud semantic segmentation is highly challenging. Some existing work ignores the loss of crucial information caused by sampling and projecting. Others use modules with high computational complexity because of the pursuit of precision, challenging to deploy in the vehicle platform with limited computing power. This paper proposes Fusedown/Fuse-up modules for efficient down-sampling/up-sampling feature extraction. The modules combine the transformer in vision integrating the global information of the feature map with the CNN extracting local feature information. Based on these two modules, we built the transformer and CNN fusion network called TCFNet for point cloud semantic segmentation. Experiments on the SemanticKITTI show that our suitable combination of transformer and CNN is necessary for semantic segmentation accuracy, and the mIoU of our model can reach 82.7% at 10 FPS. The code can be accessed at https://github.com/donkeyofking/TCFNet.git.