Object Detection And Tracking Technology For Video Sequences

As the improvement of equipment in public surveillance system, theavailable video data size becomes larger and larger. However, the massdata brings trouble and burden to the operator in the traditionalsurveillance system. The operator needs to check each frame in manyvideo files for detecting some abnormal events. This boring work makesthe surveillance system unreliable. In order to make the mass video dataeasier to analyze, the appearance of intelligent surveillance systembecomes inevitable. In our work, we put forward some methods for thetask of detecting and tracking pedestrian and car in the video sequences bythe technology of digital image processing and machine learning, whichwould make the surveillance system more reliable.In our research, we divide the surveillance system into three parts,detecting pedestrian and car, tracking objects and analyzing the trackingresult. We mainly concern about the first two parts in our work. In thedetecting part, we adopt the Gaussian Mixture Model to extract theforeground blobs and analyze each blob by their scale. For finding theaccurate position of objects, an Adaboost classifier is used. In the trackingpart, three target features are discussed. A robust tracking method based onthe cooperation of particle filter tracker and boosting learning algorithm isproposed.For addressing the problem of tracking non-rigid object, we propose ablock tracking framework. The problem of single object tracking is takenas a problem of tracking many sub-blocks of the object. The blocks of theobject are selected by the condition number method. We build a startopology model to connect each block. In tracking process, we track each block respectively, then estimate the center of the object according to theirgeometry relationship.The proposed methods in our work are realized by C++programmingand tested with several video sequences. The expected result is achieved inour experiment.