Semantic segmentation of 3D point cloud based on self-attention feature fusion group convolutional neural network

The existing algorithms ignore the profound relationship between global single point features and local geometric features. This results in the lack of discriminative captured local geometric information and increases the difficulty of effectively identifying complex shape categories. This paper proposes a semantic segmentation algorithm for three-dimensional point clouds based on a self-attention feature fusion group convolutional neural network. First, the proxy point graph convolution of lightweight network is designed to extract the local geometric features of the point cloud. Then, the group convolution operation is added to reduce the amount of calculation and complexity and enhance the richness of features with less redundant information. Second, the feature information exchange between different branches is carried out through the Transformer module to ensure mutual compensation between the global and local geometric features and to enhance the completeness of features. Then, the underlying semantic features of the point cloud are fused with the original point cloud to expand the local neighborhood perception field and obtain high-level context semantic information. Finally, the features are input into the segmentation module to complete fine-grained semantic segmentation. The experimental results show that the segmentation accuracy reaches 79.3% and 56.6% in the S3DIS and SemanticKITTI datasets, respectively. This algorithm can extract the key feature information from a 3D point cloud using fewer network parameters and exhibits high robustness of semantic segmentation. © 2022, Science Press. All right reserved.