Visual Tracking In The Surveillant Camera System

Today in the city visual surveillant system plays an important role for thepublic security. Numerous cameras and huge video data storage lead to theenthusiastic research on the intelligent video analysis. In the literature of intelli-gent video analysis, visual tracking, especially the human tracking is one of thebasic and challenging research topics. In this thesis, we frst discuss the develop-ment and trend of the visual surveillant systems. And then focus on the visualtracking literature. Visual tracking frstly detects interesting objects from thesource video sequences and efectively tracks them later. Based on the trajectoryanalysis, tracker is able to predict their behavior. In this thesis, we discuss indetail the visual tracking application in the camera surveillant background. Allthe experimental video sequences are taken from daily life. The tracking objectsare chosen to be human beings in most cases. Stemming from the probabilisticinferencing and directly local template matching, which are components of thetwo categories of visual tracking respectively, we develop diferent visual track-ers for diferent applications. All the ideas, techniques and contributions aresummarized as follows:At frst, chapter1overviews the development of visual surveillant systemand visual tracking technology.Secondly, chapter2proposes the visual tracker which combines the detectingand fltering together. By designing an interface between detecting and tracking,detector is guided to update the reference and helps to improve the tracker’sperformance. The proposed method is proofed to be efective in diferent testingscenes.Following chapter3designs a framework of local feature tracking techniquebased on the grid and interest point. With respect to the application, cascade feature model is discussed and implemented. The proposed local feature basedmethod is used for relay tracking in the camera networks.Then chapter4develops a region fnding method and a spatially topologicalcolor histogram for appearance modeling. Considering the bandwidth limitedwireless camera networks, we propose a temporal and regional covariance matrix.This matrix model keeps a low dimension in size and performs well in the wirelessplatform.Finally, Chapter5summarizes the contributions of the thesis and discussesthe perspectives of future research.