Research On Video Moving Object Tracking Algorithms Based On Binary Classification

Video moving object tracking technology is a hot research topic in computer visionnowadays. It is widely used in video object classification and behavior analysis, videocompression, human-computer interaction and it has important significance and boardapplication prospects. In recent years, video moving object tracking technology has developedrapidly. However, cluster background, illumination variation, occlusion and other factors inreal application scenarios often affect the tracking results seriously, and even cause trackingfailure. The robustness of tracking as well as the accuracy and real-time requirements are keyand difficult problems which remain to be solved. This thesis focuses on object trackingtechnology based on binary classification. The main work is as follows:Firstly, this thesis analyzes the advantages and disadvantages of different video objecttracking algorithms and their application scenes, focusing on the object tracking algorithmswhich treat the tracking problem as a binary classification problem with online self learning.The basic framework for binary classification based tracking algorithms and sample updatestrategies and other issues are analyzed.Secondly, two semi-supervised learning methods used in tracking algorithms based onbinary classification are introduced. Also theory of support vector machine and its onlinelearning problems are discussed.Thirdly, a tracking algorithm based on binary classification is proposed. The proposedalgorithm extracts the positive and negative samples from the first frame to initialize thesupport vector machine classifier. Then the samples in the subsequent images are classified bythe support vector machine classifier, and the classifier is updated online to adapt to meet thechange of object appearance. Multi-scale Haar-like features are used to represent the target.The proposed tracking algorithm does not make any prediction on trends of object movement.It is suitable for object tracking with fixed cameras and moving cameras. Also it can handlesevere occlusion problem for a short time, and it is adapted to meet illumination variation andappearance variance caused by change of their shape. Compared to a few representative tracking algorithms proposed in recent years, the proposed algorithm shows its superiority onthe dataset.