Multiple Target Tracking With Inference In Structural Model

With the development of video camera, the video tracking plays animportant role in our daily life. The target tracking in video sequences is a verychallenge topic, because there are many unstable factors in tracking. Forexample, the tracking algorithm will be influenced if the background is similarto target. Additional, obviously occlusion and appearance changing mean thatthe algorithm will lose the target. For multiple target tracking, except theproblems mentioned, the interaction among targets also affects theperformance of tracking algorithm.My research target is to build a novel structural model to solve theocclusion and appearance changing problem in multiple target tracking andimprove the tracking accuracy. In details, the structural model contains theinformation of detector, tracker and spatiotemporal context. The optimalsolution of the model can represent the tracking result.The main contribution of this thesis is that a novel structural model isproposed, which fuses the result of detector, tracker and the spatiotemporalcontext. Additional, in order to improve the robustness of detector, I proposed a strategy to combine the results from on-line detector and off-line detector.In the proposed structural model, every frame will be processed byon-line detector and tracker to get targets’ candidates. Then the off-linedetector is utilized to evaluate the selected candidates, which will be inputtedto data association function. The data association function contains two factors，the evaluation scores of candidates and the spatiotemporal context informationamong targets. The optimal solution of the function is the final tracking result.I utilize video sequences I captured, from CAVIRDATA and from PETS.The experimental results show that the occlusion and appearance changingproblem can be solved during the tracking process with proposed trackingmodel. In addition, the proposed tracking model can be applied to analyze thestructural information within a single target or the relationships amongmultiple targets.