The Research Of Human Detection And Tracking Algorithm In Surveillance Video

Intelligent video surveillance is an emerging research in the field of computer vision, and commonly used in the field of security, transportation, intelligent vehicles.Human detection and tracking algorithm is a core content of intelligent video surveillance system, so the study of relevant althogrithms is of great significance for the improvement to the system performance. In order to achieve accurate, real-time pedestrian detection and tracking in monitoring scenarios, this paper deeply research and analysis the moving object detection algorithm, human detection algorithm and pedestrian tracking algorithm, and then the algorithm is improved to complete a real-time pedestrian detection and tracking system with good accuracy. For the above system, this paper completed the following work:First, we design and implemente the moving target detection based on background subtraction algorithm. We use multi-mode mean model as a background model to complete the foreground extraction. Then for the extraction of binary image we present a relevance filter mask image filtering method. Finally, we use three scans connected component labeling algorithm to extract the connected forground. In order to accelerate the speed of modeling and reducing the calculation, we downsample the video image in background modeling.Then, we use a multilevel head-shoulder HOG(histogram oriented gradient) feature and cascade adaboost classifier to detect pedestrian in the moving region. In order to solve the occlusion problem between pedestrians,we use the HOG features of head and shoulder area instead of the whole pedestrian; In order to accelerate the detection speed without affecting detecting accuracy, we optimizatite HOG the process of feature extraction by the integral histogram; To solve the problem of uneven samples, we constructe a cascade human classifier; In order to solve the case of multi-scale pedestrian detection,we present an image pyramid approach to search people in multilevels, and then optimily thre layers of the pyramid, the target region integration methods.Finally, we use a feature template matching approach to implemente pedestrian tracking. The tracking template integrate of the color histogram feature, LTP texture and gradient histogram feature,and these characteristics were analyzed to reduce the dimension of features.In order to reduce the search template area, we combine moving region and motion predicting result to speed up the search.The pedestrian detection and tracking algorithms presented by this paper were experimented in video surveillance, shows that it has good accuracy and timeliness.