Created by March from Southwest University of Science and Technology.

In recent years, with the rapid development of technologies such as intelligent robot and unmanned driving, higher and higher requirements have been put forward for the understanding of scenes of 3D point clouds. The recognition of point cloud objects, as an important part of the high-level semantic understanding of point cloud data, occupies a very important position in the understanding of 3D scenes. The deep learning method can fit higher dimensional functions and obtain more critical feature information, which has better generalization ability and robustness for features such as disorder and sparsity of point cloud data. This paper mainly introduces the following three types of point cloud object recognition algorithms based on deep learning: voxel-based method, multi-view-based method and original point cloud data-based method. At the same time, based on PointNet and PointNet++ which process the original point cloud data directly,， the improved algorithm PointNet S was successfully implemented after continuous thinking and experimental verification. PointNet S makes full use of the characteristics that local points often have similar geometric structures. It uses KNN search to determine local areas, and then extracts local features of objects, and successfully integrates with global features. The experimental results show that the overall classification accuracy of PointNet S on the ModelNet 40 dataset is higher than the original PointNet and PointNet++ algorithms by 2.53% and 0.98%, respectively. After the number of input points is randomly lost by 80%, the accuracy rate only decreases. 2.31%, significantly better than PointNet's 4.9% and PointNet++'s 10.01%. This shows that PointNet S not only achieves better object recognition, but also enhances the robustness of the network.