Research On Wavelet Feature Based On Genetic Particle Filter Tracking Algorithm

Currently to solve non-linear non-Gaussian state estimation methods are mainly extended Kalman filter and non-marginal Kalman filtering in the process of target tracking systems. The former uses Taylor expansion to nonlinear system linearization, then apply the Kalman filter method to estimate the system state, it is due to the presence of higher-order truncation error caused by filtering accuracy is not high or even diverge; the latter change in the basis of sight, the use of Gaussian densities can embody the true mean and covariance of sampling points of the state vector of the posterior density function approximation, can be a posteriori mean and covariance of state estimation accuracy to the second-order. Another for non-linear non-Gaussian estimation problem is the particle filter, which is a recursive Monte Carlo method, by random sampling to approximate the probability density function. If the dynamic model to meet the Kalman filter assumptions, Kalman filter is optimal, the particle filter is the second-best. In a typical situation, the tested systems are often unable to meet the optimal situation, this time particle filter is usually better than the other sub-optimal filtering method.The current particle filter itself, there are still many problems, such as how to choose the proposal distribution, particle degradation, loss of particle diversity issues and how to achieve the particle filter algorithm parallel processing issues such as particle filter algorithm in-depth study of great significance to. Particle filter algorithm for the problems, this author made the following major tasks:(1) Study the traditional particle filter algorithm, mastered the principle of particle filter, mainly including Bayesian importance sampling (BIS), sequence of importance sampling (SIS), an important function, re-sampling, analysis of the degradation and particle collection particle filter algorithm for the overall understanding of the process.(2) In order to solve the particle filter particles appear in the iterative process degradation and deprivation, this paper genetic algorithm is introduced into the particle filter algorithm, that is, re-sampling process in sampling and the introduction of crossover and mutation operations.(3) As the Gabor wavelet features of the process of describing the objectives for the geometric distortion, brightness change, and the noise is not sensitive to in the video tracking for the moving target there is a big scale, rotation, affine distortion, brightness, contrast changes, and the target there is partial occlusion Under such circumstances, be able to achieve a stable target tracking. Therefore, this paper wavelet transform theory was introduced into particle filter algorithm, that is, to build target Gabor Wavelet Network (GWN), said (that is, construct wavelet feature template), proposed particle filter based on wavelet feature tracking.(4) Genetic algorithm and wavelet transform for the advantages and characteristics of this genetic algorithm and wavelet transform theory, combining the same time introduced into the particle filter algorithm to achieve the objectives in different environments stable tracking.