Pedestrian Tracking And Identifying In Video Surveillance

Pedestrian tracking and identifying are significant components of smart video surveillance system. In order to extract interested information from vast surveillance video data, pedestrian detection and tracking in the video are required; also on the purpose of searching specific target in the surveillance video, pedestrian identifying and matching should be applied in the video sequence over different cameras. Based on the summary of previous algorithms, this paper proposed a video pedestrian detection algorithm integrating Histogram of Oriented Gradients and Gaussian Mixture Model, and with the usage of the result of pedestrian detector and its classifier's confidence output, this paper improved the observation model of the traditional particle filter by integrating them and color information, achieving correctly pedestrian tracking result. This paper introduced a pedestrian appearance model by fusion of multiple features, integrating spatial-temporal information between cameras, and successfully realized pedestrian identifying and searching. The primary work of this paper contains:(1) With regards to the pedestrian detection algorithm on the basis of Histogram of Oriented Gradients in static image, we introduced the integral image to speed-up the process of statistics of oriented gradients. Simultaneously, to solve the problem of redundant computation in the region with no obvious feature change during two consecutive frames, we adopted the GMM model to build a statistical feature model, and decide the unobvious region based on it, thus reduced the detection time in the video.(2) Under the framework of particle filter, this paper realized a pedestrian tracking algorithms by integrating the result of pedestrian detection and the classifier confidence of the detector into the object observation model, and also combining the traditional color feature.(3) This paper realized an effective pedestrian identifying and matching algorithm via building the appearance model by fusion of weighted HSV color histogram, color ratio of upper and lower body and MSCR features, compensating the feature variance over cameras and integrating the spatial-temporal topology relationship.