Research On Target Recognition Method Based On Vehicle-borne LiDAR

With the development of economy and technology,the number of mileage of highway and cars is also increasing.While facilitating people's travel,it also brings many problems.One of the most prominent problems is traffic congestion and accidents.In order to solve these problems,driverless cars emerge as the times require.In the driverless system,the most important is the environmental sense system,which directly determines the safety of the driverless vehicle.Traditional environmental sense system uses image sensor as its main sensor to collect environmental information,however,under the condition of poor light environment,its disadvantages are also highlighted.The appearance of vehicle-borne LiDAR makes up for the deficiency of image sensor.The appearance of vehicle-borne LiDAR compensates for the deficiency of image sensor.It can work normally in dark light and can fully perceive the spatial information of the environment.In recent years,the application of vehicle-borne LiDAR in environmental perception has become more and more widespread.Usually,vehicle-borne LiDAR is only used to construct high-precision maps in driverless systems,in order to make the vehicle-borne LiDAR competent for the task of environmental awareness,this paper designs a method of using the vehicle-borne LiDAR to perceive its surrounding environment and recognize the human and vehicle targets.The main contents of this paper are as follows:(1)Firstly,the working principle of the vehicle-borne LiDAR is introduced.The LiDAR scans the environment and obtains the original laser point cloud data.The three-dimensional coordinates of the laser data points are obtained by visual analysis of the point cloud data through ROS and PCL.After that,the characteristics of laser point cloud data are analyzed,and the processing methods and procedures for subsequent target recognition are developed.(2)The original laser point cloud data are pre-processed,including laser point cloud data down-sampling,ground point cloud data segmentation and laser point cloud data clustering.In the process of data reduction,Voxel Grid filter is used to reduce the number of point clouds and the operation load of the system.After ground point cloud segmentation,the irrelevant point cloud will not enter the clustering operation,which improves the accuracy of clustering operation.In the operation of laser point cloud clustering,Euclidean clustering algorithm is selected by comparing the principles of several point cloud clustering algorithms.A series of point cloud clusters are obtained by clustering the point cloud data which are pretreated in the first two steps,which provides candidate targets for subsequent target recognition operations.The above three steps of pre-process operation have been verified in the experiment,and the experimental results are good.(3)The main process of human and vehicle target recognition is to identify vehicle and non-vehicle targets first,and then pedestrians in non-vehicle targets.Vehicle recognition process is: point cloud target clustering;extracting point cloud eigenvalues;using support vector machine(SVM)training classifier to recognize vehicle targets and pedestrian targets.This chapter describes in detail the features used in vehicle and pedestrian target recognition.In the problem of vehicle target recognition,considering the occlusion between objects and the uncertainty of vehicle external shape,two special feature descriptions are found based on the characteristics of laser point cloud data : distribution characteristics of reflection intensity and height profile of objects.In the problem of pedestrian target recognition,the current research status of pedestrian detection method based on laser point cloud data is analyzed.Considering the performance of equipment and the efficiency of algorithm,the geometric features,i.e.the ratio of length to width of point cloud target,are selected to judge the pedestrian target quickly,which ensures the efficiency of target recognition.In this paper,the laser point cloud data acquisition,laser point cloud data pre-processing and extraction of target feature information to be identified to train SVM classifier and other operations,a better realization of the laser point cloud data of vehicle recognition.