Particle Filter Based Parameter Estimation And Image Tracking Algorithm Study

Particle filter method and image tracking technology both are the current hot studied and applied topics. This thesis mainly focus on particle filter based methods, study the application in uncertain parameter system and image tracking, to improve the reliability and robustness of the method when applied in these fields. The main work of this thesis contains:Firstly, the thesis reviews the development of bayesian filter theory and the evolution of image tracking technology, and then introduces the fundamental principle of particle filter method. Through simulation experiments, the thesis demonstrates basic applications of particle filter method in state estimation and image tracking.Secondly, to solve the problem that particle filter based parameter estimation method is easy convergent to local optimal values, the thesis proposes a particle filter based likelihood weighted parameter estimation method. Based on the traditional particle filter method that estimates state changes, the thesis iteratively updates the estimation value of parameter through EM method. To deal with the problem of slow convergence and easy convergent to local optimal values in online EM algorithm, the thesis combines the current parameter estimation into calculating the updated likelihood and updates the step size dynamically. The simulation result shows that the algorithm can be quickly convergent to the true value. By using dynamic parameters in the motion model of the image tracking algorithm, and making use of the proposed algorithm in the thesis to estimate parameter values, the problem that the algorithm can not produce effective prediction particles due to complex object motion in fixed motion model is solved very well.Thirdly, to improve the robustness of image tracking in complex scenes, the thesis proposes a mixed measure model for particle filter based image tracking algorithm with SIFT feature. To avoid the distraction problem of background noise when only single color histogram information is applied, the SIFT feature is combined with color histogram information. Meanwhile the algorithm avoids the problem that when there're no enough SIFT feature matches due to rapid change of target, image tracking algorithm with single SIFT cue is impossible to continue. The mixed features of SIFT and color histogram, achieve complementation of information, improve robustness of image tracking algorithm. Experimental results show that the algorithm can track the image object even when the object undergoes complex motion in the scene that similar distraction objects and occlusion are possible exist.