Research On The Method Of Feature Extraction And Recognition Of Point Clouds From LiDAR

Feature descriptors are the core of 3D point cloud data processing,and play an important role in object recognition,3D reconstruction,geomorphic reconstruction,positioning and other tasks,which is also the focus of current research.Feature descriptors construct feature vectors based on spatial and geometric information contained in local surfaces of key points according to certain rules.A large number of existing descriptors are obtained by combining different feature vectors to improve the accuracy.These descriptors which fuse multiple feature vectors not only improve the accuracy,but also increase the dimension of feature descriptors and reduce the operation efficiency sharply.Along with the further development of artificial intelligence,the application of automatic driving,robots and other intelligent equipment in human production and life also puts forward more requirements on the real-time performance of algorithms.In order to improve the operation efficiency of descriptors and reduce the memory overhead,this paper focuses on the following research:(1)The theoretical basis of 3D object recognition based on local features is comprehensively analyzed.The common methods of key point detection are introduced in detail and reproduced.The specific methods and steps of feature description and descriptor evaluation are described from three aspects: local reference frame establishment,local feature description and feature descriptor evaluation method.The common methods of distance measurement criteria are discussed.Otherwise,The construction method of standard data set of point cloud commonly used in target recognition is introduced and some models are visualized.(2)In order to solve the problem of low computational efficiency caused by single neighborhood selection method and large descriptor redundancy,a new hemispherical neighborhood selection method is proposed in this paper.In this method,the local surface of the point cloud object is accurately divided by the tangent plane,so that the selected hemispheric neighborhood can retain the neighborhood information equivalent to that of the spherical neighborhood,while reducing the redundant grids caused by a large number of spherical neighborhoods.The time of calculating descriptor and target recognition and the memory required to store feature descriptor are reduced while the discriminant ability of descriptor is guaranteed.(3)Aiming at the problems of floating-point feature descriptors,such as slow feature matching,long target recognition time and high memory overhead,this paper designs a new binary encoding method to encode existing floating-point feature descriptors,and obtains a binary feature descriptor with higher operation efficiency.This use of the external threshold module and internal threshold module maximum retained the floating-point characteristic differences between adjacent dimension characteristic,making the binarization features remained with floating point performance,at the same time greatly reduce the memory overhead of descriptor computation and storage and feature matching,target recognition.