Recognition And Motion Information Extraction Of Pedestrian And Vehicle Using Laser Data

In recent years, recognition and tracking of pedestrian and vehicle arises in avariety of different research fields as a basic research question, including intelligentcar, intelligent transportation, crowd behavior analysis, traffic flow analysis, andmany others. The technology of pedestrian and vehicle recognition and tracking isgradually used in some practical application fields, such as safety driving assistsystem, autonomous vehicle and others. The research of pedestrian and vehiclerecognition and tracking has important research value and practical significance.In research of target recognition and tracking, the vast majority of researcherswork with video data. Due to the limitation of the video data which is got from theimaging planes, it is lack of target’s depth information. It cannot reflect the target’sposition and motion information in three-dimensional space. The method that getstarget’s information in three-dimensional space by combined video data frommulti-angle video, requires complex calculation. And there are some calculationerrors in these information. It cannot obtain target’s spatial information quickly andaccurately. It is difficult to meet the requirements of the practical application in termsof timeliness and accuracy.3D laser scanning technology can reflect the target’sposition information in three-dimensional space. It introduces a new data acquisitiontool and a new idea to the problem of pedestrian and vehicle recognition and tracking.The research of pedestrian and vehicle recognition, tracking, and motion informationextraction using laser data are significant in the field of scientific research andpractical application. This thesis focuses on the problem of pedestrian and vehiclesrecognition and motion information extraction using laser data. And a feasiblesolution is proposed in this thesis for the problem. The main contents include thefollowing aspects:1) Analyze the characteristics of laser data and select initial feature collection oflaser data. The characteristics of target’s laser data are analyzed in the geometry andstatistics. Target’s initial feature collection which reflect the trait of target issummarized from the following aspects: the intensity of reflected laser point, spatial dimension, level distribution of laser data, and discrete degree of laser data. Theseanalyses can put forward a method how to select eigenvalues and identify objectsusing laser data.2) Put forward a method of recognizing pedestrian and vehicle. The processwhich identifies pedestrian and vehicle target using laser data is divide into two steps:classify the laser data roughly and train a classifier to identify objects using theAdaboost algorithm. In the course of the rough classification, laser data is dividedinto two categories including the non-target category and the target category usingsegmentation method based on projection plane mesh. Then the number of laser datawhich needed to be processed for the problem is reduction. The target is clustered toselect the target candidate region using clustering method based on multi-scaledistances. In the process of object identification, the multi-class target recognitionproblem is split in multiple single-class recognition problem. In the process oftraining classifier, the features of laser data in statistical and geometric is susceptibleto environmental change. It is difficult to obtain a high-performance classifier directly.So this thesis train a classification by Adaboost algorithm based on the idea ofintegrated learning.3) Put forward a method of extracting motion information of pedestrian andvehicle using laser data. In the process of extracting motion information using laserdata, the trajectory information of pedestrian and vehicle are extracted usingmulti-target tracking method based on particle filter algorithm. In the method, eachtarget is tracked using the particle filter algorithm separately. A conflict judgmentmodel is used to determine and deal with the conflict. In this way, the particle filteralgorithm commonly used in video tracking algorithm is used for target trackingproblem using three-dimensional laser data. And then, target’s acceleration, speed,average speed, the accumulated motion direction, the distance and angle amongtargets are calculated in this thesis by using the position and trajectory data obtainedfrom the algorithm of recognition and tracking. These values will represent target’ssports trend and interactive sports trend among targets.