Research On Human Target Recognition In Complex Scene Based On 3D Lidar Point Cloud

Target recognition is an important research direction in the field of computer vision.Target recognition often be used to distinguish one or a class of special targets from other targets.Usually,target recognition is a typical task in pattern recognition system,generally include four major steps:data preprocessing,feature extraction,classifier design and training,pattern classification.Currently,many researchers are working continually on target recognition with two-dimensional image.With the development of three-dimensional imaging technology,the techniques of structured light measurement,laser scanner and TOF has been much riper.Therefore,the stereo coordinates of 3D surface can be obtained accurately and rapidly.Thus,we can generate 3D scene data,perceive and understand the surrounding better.Therefore,this topic aims to solve the problem of human target recognition in complex scene based on LIDAR point cloud.For this reason,we established a stable and efficient 3D LIDAR scanning system.In addition,this topic designed a target recognition scheme based on complex point cloud scene,all those can run on the system we have established.This paper focuses on the following contents:Firstly,the combination of the 3D scanning LIDAR system.Aiming at the features of LIDAR scanning,this paper presented the system hardware and software design,and its scan method.Then,a stable and rapid scanning system was completed.Secondly,pre-process of the scene point cloud.Specifically,this section covers the following steps.According to the least-squares method,it achieved the ground points extracting by the planar fitting.A new background point removal method based on the orderly point cloud,which was used to remove the background points in complex scene,has been proposed.Then,coordinate system transformation and ground point removal operations were performed for scanned point clouds.Also,two types of noise point removal method were compared and analyzed,finally the radius filter was determined.Thirdly,human target recognition.This section compared and analyzed several scene point cloud segmentation schemes,and proposed a new segmentation method based on a projection method combining the advantages and of other schemes,which improves the accuracy of target extraction.Then,this section compared and analyzed the performance of the three feature extraction methods for point clouds,and finally adopt the point cloud feature extraction method based on ESF.Also,we determined the classification model based on support vector machine(SVM).Finally,through the above steps,a high availability human target recognition scheme in complex scenes has been realized.In the end,research on point cloud processing scheme based on deep learning.This section analyzed two point cloud processing algorithm models based on deep learning,PointSIFT and SO-NET.We introduced the principle of the network and its performance on point cloud data.This part is used to guide the follow-up research work.