Research On Video Object Tracking With Kalman Filter

In the application and prospect of computer vision technique, video target recognition and tracking was an important research subject. It had got extensive applications in such fields as human-computer interaction, video monitoring, intelligent buildings, and medical image processing. After large numbers of researches by many scholars at home and abroad, the video tracking technology has made the great development of multi-direction, many excellent tracking algorithms are proposed. However, these algorithms are often based on some special conditions, such as in the static background, indoors, no occlusion, etc. However, how to track target in video sequences robustly and to intelligent accurately analyze the implied information of the moving target and to achieve a goal of automatically detecting and tracking system are still challenging task.In this paper, we firstly discusses the background knowledge, significance and current situation of the research, and summarized the fundamental methods of the video target tracking, including common tracking algorithms, trace flow, main difficulty and technical requirement, etc. Then we reviewed kalman filter theory and represented primary algorithm and property of kalman filtering from the bayesian filter theory. Finally, the extended kalman filter and unscented kalman filter were discussed.In this paper, the theory of detecting and dividing moving target was proposed, we also contrasted several detecting algorithms of moving target and then explored the three frames difference detailedly. Finally, we proposed a semi-automatic target dividing method. The simulation results showed that we could get moving target by dividing and extracting features using the proposed algorithms.This paper also discusses the classical mean shift algorithm, we explained the basic principle of the mean shift algorithm with the aspects of the target model, the candidate model, similarity measure, targeting location, then we analyzed the advantage and disadvantage of the algorithm and done the simulation finally.In this paper, for the lack of mean shift algorithm, we presented a video moving target tracking algorithm which compromised the modified mean shift and self-adaptive kalman filter. We used frame difference and region growing method to divide target and get main color information. During the tracking course, the self-adaptive kalman filter was used to estimate the initial iteration position of every frame. Then we used improved mean shift to get tracking position which acting as measured value feed back to the self-adaptive kalman filter. Furthermore, we introduced screen ratio factor to dynamic accommodate parameter of kalman filter. Accordingly, the modified mean shift algorithm could calculate the succeed state of the target. We could track the target accurately even if it was sheltered at short time. The results demonstrated that the algorithm has great improvement relative to mean shift, it could detect and track the video moving target continuously and had preferable robustness for occlusion.