The Application Of The Improved Particle Filter On Visual Object Tracking Algorithm

Visual object tracking is a key issue in computer vision and fuses the theory and knowledge of different fields such as image processing, pattern recognition and artificial intelligence, which is widely used in the field of intelligent control, intelligent transportation, image compression, human-computer interaction, medical diagnosis etc. Object tracking can be seen as a non-linear and non-Gauss system state estimation problem. Particle filter algorithm is the most effective method to solve this problem. However, particle filter in the current is not mature enough and still exists many drawbacks. This article focuses on the study of particle filter, and proposes several methods as follows:(1)Two improved UPF algorithm are proposed.Considering that proposal distribution can not fully use of the measured information and particles degenerate in the standard particle filter, the research studies from the point of the sigma-point sampling methods of unscented transformation:The first algorithm uses Gaussian distribution which has the same mean and variance as the particles to sample particles, and the sample particles are generated through distribution function based on UKF algorithm of Gaussian sigma-point sampling, then uses the Metropolis-Hastings to optimize particles, and the state of estimation is achieved by the weighted summation of particles. The second algorithm uses minimal skew simplex sampling strategy to sample particles, and improves the accuracy thtough the scale corrrection. Then adopts iterated Kalman filter to optimize the mean and variance of the state which are obtained in the prediction stage,and generates the sample particles through the achieved distribution function. Finally, it uses the Metropolis-Hastings to optimize particles and estimate the state. Simulation results show that both of the above improved algorithms reduce particle degradation existing in the particle filter, and improve tracking accuracy.(2)An improved UPF tracking algorithm based on multi-feature fusion is proposed. To solve the robustness problem and poor use of the latest measurement information in the object tracking with single feature. Firstly, the algorithm is improved by using UPF algorithm with scaled minimal skew simplex sampling strategy and IKF algorithm, then adopts uncertain measure method to fuse the color and texture features of the object and tracks the object with the framework of the improved algorithm. Simulation results show that the proposed algorithm improves the tracking accuracy, has a better effect on tracking the object under complex scenes accurately and tracks the occluded object effectively. (3)An improved particle filter object tracking algorithm based on genetic is proposed in this paper. For heavy sample depletion of partecle filter caused by the resampling process, the algorithm uses two selection operations to select and copy of the particle, to grouping ceoss breeding of particles for a particular sort of crossover probability, mutation breeding pins through the diversity of particle, in order to reduce the phenomenon of depletion of the sample particles. Experimental results show that the proposed algorithm improves the tracking accuracy, and can track object in complex condition robustly.