Detection And Tracking Based On The Behavior Of The Machine Vision Technology Research

Pedestrian detection and tracking based on Machine visual has been applied verywidely in the fields of video surveillance,3-D reconstruction, and autonomous vehiclenavigation, etc. Thus, it is of great significance to study pedestrian detection andtracking based on machine vision. Because the human body is non-rigid, and there area lot of interactions among the pedestrians or between the pedestrians and scenes, andthe influence of motion of the camera, the pedestrian detection and tracking faces themost difficult challenges among those tasks mentioned above. Technologies ofpedestrian detection and tracking are applied gradually in clutter scenes, but they arestill in the initial stage of rapid development.Several key technologies of detection and tracking for pedestrian are studied inmy thesis as follows:In the first chapter, the application background, the current situation of researchand the content of my thesis are expounded.In the second chapter, the basic theoretical knowledge of the machinevision-based pedestrian detection, the advantages and disadvantages of the methodscommonly used to describe image feature are expounded, and the mean-shiftcorrelation algorithm is descripted, too.In the third chapter, to solve the problem of the dense zoning issues in thetraditional Shape Context method of pedestrian detection, we proposed an improvedmethod (ISC) by changing the zone and also use the method of the histogram intervalfuzzy. My ISC has some tolerance for deformation, and enhances the ability ofanti-interference to the debris for the background. Experiments show that ISC is moredesirable.In the fourth chapter, pedestrians in the more complex environment arevulnerable to changes in illumination, pose, occlusion, and other unfavorable factors,we used the random forest classifier and pair-compare feature to detect pedestrian.This statistical learning method based on a large number of samples can deal with to some extent the problem of illumination changes, blocking and the attitude changes.Experiments show the effectiveness of my algorithm proposed.In the fifth chapter, an efficient and robust pedestrian tracking algorithm isproposed. The tracking algorithm based on the original mean shift algorithm has alower tracking rate or lost the object, when another object is in front of the object, orthere are background disturbances. To solve this problem, the improved algorithmbased on weights given to the object and the background is proposed. This algorithmcan reduce the relevance of background information and improve the targetlocalization. Also, real-time updating of the target template is provided to ensurestable,real-time tracking of object in videos. Experimental results show that theproposed method is more robust to present object and the effect of tracking isimproved obviously.In the sixth chapter, we summarized the main content and innovative points ofour thesis. And the related research in the future of pedestrian detecting and trackingis prospected.