LiDAR Point Cloud Semantic Segmentation Method Based on Multi-scale Contextual Feature

Semantic segmentation of point clouds scanned by LiDAR is one of the means for robots to perceive the environment autonomously. Aiming at the sparse and unstructured characteristics of LiDAR point clouds, we use the spherical projection formula to project LiDAR point clouds to a dense range image. A 2D convolutional neural network based on the encoder-decoder structure is used to perform semantic segmentation on the range image. After segmentation on the range image, we re-project the semantic result of the range image to the LiDAR point clouds using a kNN method. To extract the context features of the range image, we design a multi-scale contextual feature extraction module based on the feature pyramid network, so the encoder-decoder network can better obtain the semantic features of the range image. The experimental results show that the mIoU of the proposed model is 55.2% and 45.0% in SemanticKITTI and SemanticPOSS, which is 3.0% and 16.1% higher than that of the RangeNet++ network, respectively.