Target Detection And Tracking Based On Haikang Video Monitoring System

With the rapid improvement of people’s life, industrial production, home life servicesand other public places all require intelligent video surveillance system to assist, and theperformance of intelligent monitoring system is be demanded more and more high. Theso-called intelligent video surveillance system is in no need of manual intervention, by usingdigital image processing, machine vision, pattern recognition and other related technologiesto processing the video information we have acquired, including detecting the interestedtarget, tracking the interested target, target action recognition and behavior analysis in thevideo. Currently the monitoring system on the market has been put into use mostly don’t haveintelligent analysis ability or not intelligent enough, so under the premise of withoutincreasing the hardware cost and greatly reduce manpower cost, and strive to enhance theperformance of the existing video surveillance system, try to reach the advanced monitoringsystem functions, is a very important research topic.This paper is mainly aimed at the Haikang intelligent video surveillance system existingin our laboratory, its intelligent analysis functions is very few now, my task is to add targetdetection and tracking algorithms on obtained real-time Haikang video streams, andeffectively extract the object features of the target of interest, such as the target contour andtrajectory, so we can found a solid foundation for the action recognition and behavior analysis.And the ultimate goal of us is to make the performance of the Haikang video monitoringsystem to reach to the existing advanced intelligent video surveillance system.Firstly using the relevant interface functions provided in Haikang SDK to obtain thereal-time video stream transferring by the network, and change the video format from YV12to RGB format which is commonly used. Then we begin to model background method usingthe combined information of spatial and temporal information of the pixels. Using the formatN frames of Haikang real-time video to establish the codebook background model which iscomposed of a few of codewords for each pixel,this model adequately describes the timeinformation of pixel sequence; also used the same N frames to establish texture backgroundmodel according to the scale invariant local ternary pattern, for each pixel in the currentframe, if and only if it satisfies the codebook and texture background model meanwhile, it is the background, otherwise will be judged as the foreground. The codebook backgroundmodel update background by adding or deleting codewords, and the texture backgroundmodel update background by using the traditional method. Finally, this paper draws thecontour and the target trajectory using the detection results.This paper looks at the target tracking problem as two categories of the target andbackground. First, according to the results of target detection or by hand to get the initialposition of the object, we extract the positive and negative samples around the position of theprevise frame, and compute their multi scale rectangle features and hog features to trainingthe Bias classifier. Then collecting some rectangle samples in the next frame around thetarget template to be detected. Finally the maximum Bias classifier results is the objectposition, thus realize the tracking from the current frame to the next frame. Simple Biasclassifier is continuously updated in the tracking process.In this paper only target detection and tracking are done to enhancing the performance ofvideo monitoring system, analysis and action recognition and behavior analysis for the targetstill need further research。...