| Target recognition is a core subject in the field of intelligence information proc-essing,is widely used in aviation and aerospace,medical and health,navigation and guidance and other fields.In recent years,with the advantage of LiDAR,has quickly become a hot research topic in both academic and industry.Compared with the tradi-tional two-dimensional image,obtaining 3D laser radar range image by LiDAR,the true 3D characterization of geometric shape and depth information of the object,with the scale and rotation invariance.And has smaller effections in the viewpoint,scale,intensity of illumination and occlusion aspects.So it is expected to make the difficult tasks in 2D images simpler.In this paper,the 3D point cloud target recognition prob-lem of ground vehicle is studied,and the following work is carried out in the aspects of the method research of target feature extraction and selection,target recognition and the development of information processor.(1)The paper mainly analyze the typical features of the existing extraction method based on projection contour feature extraction,geometric characteristics,Spin image and SHOT feature extraction.Laborating the basic principle of these methods and analyzing all kinds of characteristics of the description of the performance,and then the extracted features are given through simulation experiments.The results show that the existing 3D point cloud features has a certain scope of application,and it is difficult to accomplish the specific recognition task for features of a single type.On the basis of this,a multi feature combination description method is proposed,which is based on SHOT features,projection contour features and length,width and height features.(2)The thesis propose an attribute reduction algorithm based on neighborhood combination measure.In the neighborhood rough set theory,to measure the attribute importance combined to the uncertainty of set and knowledge from set theory and al-gebra view.First of all,defining a new measure for the uncertainty of knowledge,named as neighborhood granule measure.Then integrated with the neighborhood ap-proximation accuracy,combination of neighborhood measure is defined for measuring the importance of attribute.According to it,the thesis propose an attribute reduction algorithm based on neighborhood combination measure.The testing on the UCI data set,the results show that the algorithm is not only achieve compact reduction,but has good classification performance.(3)A 3D point cloud ground target recognition method based on neighborhood rough set theory and Support Vector Machine is proposed.Will based on neighbor-hood rough set attribute reduction method is applied to the feature selection of 3D point cloud ground target,effectively eliminate the redundant features in the charac-teristic of projection contour feature and SHOT.Using feature reduction through the structure model of Support Vector Machine classifier.The experimental results indi-cate that reduced the feature dimension is far less than the dimensionality of original features.At the same time,it can effectively reduce the feature dimension and im-prove the real-time performance,which can guarantee the classification ability of Support Vector Machine.(4)The thesis design a real-time information processing system for target recog-nition of the 3D point cloud using the architecture of FPGA and dual DSP.Based on a detailed analysis of algorithm and resource consumption of each module,has com-pleted the hardware circuit design and realizes a parallel processing system on the frame of multi DSP multi core(using the master-slaver topology structure).The test-ing of the system manifest that the designed system can identify the target type and has good real-time performance. |
