Layout Point cloud local parallel attention feature learning network with enhanced spatial features
Xingfeng Li, Yue Dong, Hua He · 2024
In The global point cloud feature extraction method lacks hierarchy and is difficult to learn features at different scales, which limits its ability to extract local features. However, for local point cloud feature processing based on progressive sampling and grouping strategy, the irregular distribution and permutation of point cloud will bring great challenges to algorithm design. This paper proposes a point cloud local parallel attention feature learning network with enhanced spatial display features, and constructs an efficient point cloud feature learning network(LPA) that can aggregate local features and global features. LPA network only uses Transformer parallel module to capture its global information during feature extraction at the top level. Based on traditional dynamic graph convolution, superficial features are learned from local graph convolution features by using the feature differences expressed in the geometric space features of the center point and neighboring points in series. Effectively improve the ability of extracting the spatial relative distribution and geometric distance feature information of point cloud objects. In a top-down approach, a globally centrally regulated feature pyramid is proposed. Compared with existing feature pyramids, the LPA network uses the deepest explicit global information to regulate frontend shallow features, which not only captures global long-distance dependence, but also efficiently gets a comprehensive and differentiated feature representation. The model was evaluated from multiple tasks. The results show that: The OA value of the model shape classification reaches 0.93, while the mIoU value of the model for parts segmentation and semantic segmentation reaches 0.853 and 0.721, respectively. The network effectively improves the accuracy of point cloud segmentation. Realize efficient semantic segmentation of point cloud.