Research On Global Feature Extraction Of Laser 3D Point Cloud Image

The 3D point cloud data obtained by lidar can directly reflect the 3D geometric information of the measured target.Usually,the 3D point cloud data is composed of low-level data structure,points,so it is difficult to describe the target effectively.Therefore,in order to extract and describe the point cloud data at a high level,this paper studies the global feature extraction method of 3D point cloud target.Firstly,the basic principle and method of converting lidar range profile into 3D point cloud are described.And two geometric feature extraction methods,a moment feature extraction method,a shape feature extraction method are introduced.The applicable conditions and existing problems of the global feature are analyzed.Secondly,aiming at the problem of how to accurately estimate the length,width and height direction of the target in order to calculate the geometric size of the target,a target attitude estimation(PEMSPNC)algorithm based on Mean Shift point normal vector classification is proposed,and the target geometric size feature is extracted on this basis.The simulation data of target range profile are generated by the lidar simulation system,and four different situations are simulated,including no noise,different ranging accuracy,presence of noise and presence of occlusion.The target geometry results calculated based on rectangle fitting,principal component analysis,OPDVA and PEMSPNC are compared.In order to further verify the practicability of each algorithm,a set of measured data are used for experiments.The results show that the accuracy and robustness of PEMSPNC algorithm are better than those of the other three methods.Finally,aiming at the problems of large dimension and information redundancy of projection contour features,the PCA dimensionality reduction algorithm is used to remove the possible correlation between projection contour feature data and reduce the dimensionality of projection contour features.By simulating the three situations of no noise,occlusion and noise,the BP neural network is used to realize the target classification and recognition experiments based on projection contour features before and after dimensionality reduction.The results show that the PCA dimensionality reduction algorithm can effectively eliminate the redundant information in projection contour features,and improve the effectiveness of feature while ensuring its classification ability.