Research On Moving Object Detection And Tracking In Video Surveillance

Intelligent video surveillance is an emerging research direction in the field of computervision. Its main goal is to describe, understand and analyze the content of the surveillancevideos by utilizing computer vision technology, image/video processing technology and ar-tificial intelligence technology. The results of the analysis can then be used to control thevideo surveillance system itself. Thus, the video surveillance system can be improved to ahigher level. The main research contents of intelligent video surveillance include: movingobject detection, object representation, object tracking, object recognition and object behav-ior analysis.This paper is committed to the research on the moving object detection and tracking inintelligent video surveillance. And the research work mainly covers the following topics:1. An improved eigenbackground modeling method for videos by recursively applyingan error compensation process to reduce the in?uence of foreground moving objectson the eigenbackground model is proposed. An adaptive threshold method is also in-troduced for background subtraction, where the threshold is determined by combininga fixed global threshold and a variable local threshold. A fast algorithm is then givenas an approximation to the proposed method by imposing and exploiting a constrainton motion consistency, leading to about 50% reduction in computations.2. A robust mean shift based object tracking method using subclass discriminant analysis(SDA) color space is presented. SDA color space is proposed which seeks to findthe color subspace for representing pixels by maximizing the distance between theforeground pixels and background pixels even if target and background have multi-model color distributions. Foreground object position is obtained based on mean shiftlocalization algorithm using a weight map generated by the features in SDA colorspace. Further, SDA color space is adaptively updated by only using"confident"targetpixels. 3. A tracking algorithm based on the multi-cue adaptive fusion has been proposed. In thismethod, the target observation is represented by multiple cues. When fusing each cue,a fusion scheme based on model representation power and model discrimination powerhas been developed. Each cue is adaptively fused during tracking, which increases thereliability of observation and improves the robustness of observation model. Whentracking targets, particle filter has been adopted, and the observation model based onthe multi-cue adaptive fusion is included in the particle filter tracking framework. Theproposed tracking algorithm can track targets robustly in complex environments.4. Since single kernel histogram model can not be used to represent target with multipleappearances in complex backgrounds, a novel mixture of kernel histograms model(MoKH) is proposed for visual tracking. Multiple weighted kernel histograms forrepresenting different appearances of the tracked object were used. The weights andthe kernel histograms are updated adaptively. The MoKH model is used in a particlefiltering tracking framework. Results from experiments with real video data showthe improved performance of the proposed algorithm when compared with that of thestandard particle filter using a single kernel histogram model.