Study Of Multi-moving Pedestrians’ Detection And Tracking

Detection and tracking of moving pedestrian in video image sequence is one of the keyresearch topics in computer vision research, and it is increasingly widely applied to the fieldof safety precautions and intelligent surveillance. There are many confounding factors invideo such as complex background environment, pedestrian attitude change, the goal ofmutual occlusion or staggered, and light or climate changes. These confounding factors makedetection and tracking of moving target become difficult issues of computer vision researchareas. Although many scholars at home and abroad conducted an extensive and in-depth studyfor the field, also proposed many solutions, but there are still many key issues which have notbeen effectively resolved, a more mature and steady techniques or methods of detection andtrack are urgently needed. Therefore, this paper carried out in-depth research in this area,especially the problem of multi-target impact and t mutual occlusion.First propose the Adaboost detection algorithm based on HOG feature and Haar featurefacing traditional detection algorithm ineffective which the light changes, shadows and otherfactors caused. This algorithm combines the Haar features and HOG feature, It ischaracterized by: The Haar features can describe the situation when its target local graysuddenly changed; The HOG feature can preferably characterize the target’s contour and edgefeatures, So it achieved better detection effect and improved the robustness of target detection.Second, this paper do a depth research on the moving pedestrian tracking, Improved thetraditional Mean-shift algorithm, it leaded into spatial histogram mean shift algorithm, andsupplied the derivation and the tracking process of the algorithm. The space the histogrammean shift tracking algorithm not only contains the exact same color histogram information,but also includes additional spatial information of each subspace. With the advantage ofstrong anti-interference and low-impact light. it can capture more richer description of thetarget.Finally, since the histogram mean-shift algorithm is easy to track failures in multi-targetscene, especially blocking or tracking targets in close proximity will lead to loss of target，Soin this paper, introduced a new kind of mean-shift algorithm which merged Fusion Kalmanfilter and block ideology based on above algorithm.The algorithm proposed that the predicting of motion information and blocktracking should be joined.With building block model for the target obscured and doingmean shift iterative search for each block，we make sure that the block unobscured is the final tracking result to solve the problem that the scale change of the target tracked and thetarget tracking failure when it has been obscured. By doing several experiments this methodhas been validated.