Video Object Tracking With Related Multi-region Based On Kalman Filter

The application of video object tracking becomes wider and wider with the development of computer software and hardware technology. However, in the process of object tracking, object occlusion always happens and makes tracking instable. Besides, the situation of similar object and background appears to make tracking drift, even leads to tracking failure. Therefore, it is urgent to solve the problems of object occlusion and tracking drift to make object tracking stable and accurate. In this dissertation, the above problems of object occlusion and tracking drift are investigated, and the main research contents and innovative points are as follows:By choosing the color histogram model as the object feature, a feature matching method of multi-region is proposed to solve the problem of object occlusion and tracking drift. The method divides the object into multiple regions and presents the relationship among the regions by an undirected graph. On the basis of Bhattacharyya coefficient of signal region feature matching, the distance factor of adjacent regions is introduced and a global similarity measurement function of multi-region is established, which makes the positions of the rest regions corrected when some regions’tracking is abnormal. As a result, the robustness and real-time of tracking is promoted.On the basis of the multi-region feature matching method, an object tracking model of related multi-region based on Kalman filer (tracking model for short) is proposed. Kalman filer can predict and correct the trace of object movement. The prediction lays a good foundation for getting accurate observation information while the correction promotes the accuracy of object tracking. Besides, simple solution of object template update and dense occlusion is presented based on the model.In order to realize the tracking model, a region sampling method of spiral queue is proposed, and a related multi-region tracking algorithm based on Kalman filter and spiral queue sampling (spiral algorithm for short) is established. Compared with the traditional uniformly-spaced region sampling, spiral queue sampling provides the recursion formula for histogram model of sampled regions. It reduces the calculation cost and makes tracking become more real-time. The experimental results show that the algorithm can better deal with the problems of object occlusion and tracking drift, which has better robustness, real-time and accuracy compared with the classical methods based on Mean Shift and particle filter.To solve the problem that observation information depend too much on the predicted state based on Kalman filter in the tracking model, the dissertation integrates Markov chain Monte Carlo (MCMC) method into the tracking model and establishes a related multi-region tracking algorithm based on Kalman filter and MCMC sampling. Based on the initial state of Markov chain predicted by Kalman filter, a Markov chain of stationary distribution of observation density is constructed. So that under the condition of a fixed length of Markov chain, the achieved observation information is more accurate, which is used for the correcting stage of Kalman filter. The experimental results show that the algorithm not only can deal with the problems of object occlusion and tracking drift, but also is more accurate and stable. However, the property of real-time is little worse.