Research On Point Cloud Classification Based Local Spatial Features

The special spatial structure information of 3D point cloud data makes it possible to further improve the accuracy of visual application.Point cloud classification is an important research direction of point cloud data processing,and it is also a challenging topic.The focus of related research is mainly on the following aspects:(1)improving the accuracy of the point cloud classification algorithm,and(2)speedup point cloud processing while maintaining accuracy.This paper mainly studies the point cloud simplification algorithm,the feature descriptor building method and the related classification algorithm,aiming at improving the speed of the point cloud classification algorithm under the condition of maintaining classification accuracy.The main work of this paper is as follows:(1)Aiming at the problem that the current point cloud simplification algorithm is not efficient and the simplification operation leads to reduced accuracy,this paper proposes a point cloud non-uniform stratification method and a layered point cloud simplification algorithn.Firstly,the spatial structure information of the point cloud is used to complete the non-uniform stratification process;then the curvature information of each layer is calculated and the unimportant points in the layer are removed according to it,so that the single layer point cloud is simplified;Finally,the simplified layers of point clouds are spliced to obtain a complete simplified point cloud.The experimental results show the effectiveness of the proposed point cloud simplification algorithm.(2)Aiming at the problem that the classical point cloud feature descriptor has high feature dimension or low computational efficiency,this paper proposes a binarized geometric feature descriptor(BGFD).Firstly,the feature importance evaluation and feature redundancy analysis are used to select the optimal geometric feature subset,and the binarization feature representation(BGFD)is obtained which has low dimension.The comparison experiment between BGFD descriptor and classical descriptors for scene target matching shows that the low-dimensional and high-efficiency feature descriptors proposed in this paper maintain a high description ability.(3)Aiming at too long training time of high-precision machine learning point cloud classification algorithm.This paper proposes a point cloud classification algorithm based on fused binarized feature description.The RoPS descriptor is improved by the binarization method to generate the binarized RoPS descriptor(B-RoPS),and the BGFD feature is spliced to form a fUsed binary feature description,which is used as the input characteristies to the point cloud classification algorithm.The experimental result shows that the point cloud classification algorithm based on the above features speedup the model training process while maintaining high classification accuracy.