Point cloud classification method based on local dynamic edge graph convolution and global information fusion

The development of automatic driving, unmanned aerial vehicle detection and virtual reality cannot be separated from the research of 3D point cloud. Existing studies either lack of local shape capture and local spatial feature description of point cloud, resulting in poor accuracy of point cloud classification network, or pursue fine and complex local feature extractors, sacrificing extraction time and memory overhead to extract classification progress, but the results show that the effect is not significantly improved. Therefore, this paper proposes a Point cloud classification method based on Local Point Dynamic Edge Graph Convolution and Global Information Fusion. Among them, the local point dynamic edge graph convolution module improves the edge graph convolution, which can improve the local extraction feature effect without deepening the local feature extractor, and reduce the time and memory overhead. The global information fusion block concatenates multi-layer local features to extract and fuse feature information to improve the classification accuracy of the network. Our network is tested on the public datasets ModelNet40 dataset and ScanObjectNN dataset. The results show that compared with the current mainstream point cloud classification algorithms, the point cloud classification accuracy of the proposed classification method is improved, and it has good robustness.