Study On Some Key Issues Of Pedestrian Detection And Tracking Based On Video In Complex Traffic Scene

Pedestrian detection，tracking and count based on video is a research focus in computervision, and it has a wide range of applications in traffic safety, real-time traffic statistics,intelligent surveillance, etc. However, because of the different light conditions of thebackground environment, the disequilibrium of the samples in scene, occlusion andinterference, pedestrian detection，tracking and count is still facing many difficulties, whichaccuracy and stability of detection method should be further improved.Some key issues of pedestrian detection and tracking are researched in this dissertation,which mainly include: the algorithm of extracts high robustness pedestrian feature, theclassification algorithm consider the problem of imbalanced samples and the diversity ofweak classifiers at the same time and the dynamic optimization method for the key parameters,high efficiency of the classifier structure, fast pedestrian tracking based on information ofcolor, texture and space, data correlation of multi pedestrian tracking and pedestrian countmethod based on pedestrian detection and multi pedestrian tracking. Specific studies asfollows:1. A bilayer difference feature extraction algorithm is proposed. The upper layer ofbilayer difference feature is constituted by improved Assembled Binary Haar (ABH) feature,which is inspired by the idea of Local Binary Pattern (LBP) and combine binary Haar featuresthrough the LBP rule. ABH feature could enhance the ability of illumination invariance andkeep high real-time performance, which is used for fast pedestrian detection and location;then, Edgelet feature is used for the lower layer of the bilayer difference feature and check thevalidity of pedestrian which is detected by ABH feature, reduce the influence of false alarmand occlusion to detection.2. A Disequilibrium Gentle Adaboost with classifiers algorithm is proposed. In order tosolve the problem of class-imbalanced in pedestrian detection, this new algorithm lead inCost-Sensitive SVM (CS-SVM) as weak classifier and use Disequilibrium Gentle AdaBoostalgorithm to compose weak classifiers into strong classifiers; in addition, through dynamicadjust the values of the kernel parameter in CS-SVM and get a series of component classifiersdifferent with each other, and measure the diversity of component classifiers, then discardpoor and similar classifiers, which could improve diversity among classifiers and theclassification performance of the algorithm. 3. Because the parameters of cost-sensitive is very important for the classificationperformance of DGACS, a dynamic optimization method about the parameters ofcost-sensitive is proposed for choosing the appropriate parameters. This method uses a newoptimization algorithm which is proposed in this dissertation (Chaotic Particle SwarmOptimization with T mutation). The best compromise about the correct classification of thepositive and negative samples is used for optimization principle, and dynamic choose theglobal optimal solution in the range of the cost-sensitive performance.4. A combination classifiers based on hierarchical tree structure is established, whichadopts the ideology of “coarse-to-fine” to detect pedestrian step by step. Among them, thecoarse combinations classifier employs a full binary tree structure. This structure is used forrejecting numerous obvious negative samples, and selecting positive samples and somehidden negative samples as candidates; the fine combinations classifier utilize a series treestructure, which used to classify more precise from candidates.5. A fast particle filter pedestrian tracking method based on color, texture andcorresponding space information. Firstly, we extract space information of object pedestrianand disintegrate it into three local regions (head, upper body and leg); next, employ theimproved texture and color information extract algorithm to get the joint texture and colorinformation from the corresponding sub-region; finally, determine the position of object bycolor-texture similarity indicator based on space division, and get the result of accuratelytrack. In consideration of the multi thread information fusion algorithm need a larger numberof particles, and reduce the computational efficiency. Therefore, a wave integral histogramalgorithm is proposed for improving arithmetic speed.6. A multi-pedestrian tracking algorithm based on k-best MHT is improved. Firstly, forexcluding part of the tracks which have low credibility from the source, we improve the wayof generate and select the objective tracks; secondly, the link of hypothesis management ismend. In order to reduce the dimensions of hypothesis matrix, measurements are groupedbefore hypothesis matrix generation, and use group as a unit to trim or merge matrixgenerations. A pedestrian counting system is set up which is based on pedestrian detection andmulti-pedestrian tracking. The system obtains the number of real-time pedestrians inmonitoring area by detect the number of pedestrians who are in the initial frame and calculatethe number of pedestrians who pass in and out monitoring area in the follow-up frames, which could get the real-time passenger flow information statistics in the detection regions.