Research On Moving Target Tracking Methods Based On Non-linear Filter And Mean Shift

Video object tracking,which has been applied successfully in augmented reality,sensing interface and video surveillance,is an importance branch of computer vision.The main idea of video object tracking is to acquire the position and motion direction of the object using the correlation of video information in time and space.Then the collected information can be used for video analysis.To insure the practicability,the ideal video tracking algorithm should be robust and real-time.But in practice,the algorithms mostly are of great complexity and it ca be a long time to proceed the tracking algorithm.So it is still much work to do in the field of video tracking.Among various object tracking algorithms,particle filter and mean shift algorithm have been paid much more attention on for their unique advantages.Particle filter is very suitable for the environment of nonlinear and non-Gaussian.It has the advantage of high accuracy and has been widely used.However,it is an algorithm based on Monte Carlo sampling.There is the problem of particle degeneracy to constrain the tracking performance.Mean shift algorithm has been applied successfully in the field of object tracking for its real-time instantaneity and robustness.However,when target moves quickly or mutation occurs in target background,its performance will decrease heavily and even it will lose efficacy.To address on these problems,the video tracking algorithms are analyzed and studied in this dissertation.The main work of this dissertation is introduced simply as follows:Firstly,a detailed description on the principles of various video tracking algorithms,including extended Kalman filter,unscented Kalman filter and particle filter.The comparative simulation experiments are conducted to analyze the character of these algorithms.The particle degeneracy problem has been analyzed.Aiming at the problem of particle impoverishment on resampling,a new resampling method is proposed.The sample interval is divided into several small parts centered by the current particle.The resampling is conducted in these small interval by Gaussian distribution and the number of samples is proportion to the weight of central particle.Simulation experiments show that the tracking performance is almost the same as that of the traditional algorithm,while the computation complexity declined a lot.The video target tracking algorithm based on Mean Shift has been successfully used in vision tracking field due to its real-time characteristic and robustness.However,when target moves quickly with random trajectory,its performance will decrease heavily and even it will lose efficacy.Aiming at the problem,the Mean Shift tracking algorithm has been developed based on the current statistical model.In the method the target's movement characteristic was modeled based on the current statistical model and it forecasted the target candidate position at first.Then the optimal target position is renewed using Mean Shift algorithm and it is fed back to tracking filter as the next measurement.Experimental results show that the algorithm possesses the qualification of less iteration number,high tracking precision,and high reliability,and has better tracking effect when coping with rapid moving target tracking.