Research On Improved Algorithms And Applications Based On Particle Filter

Aiming at the complexity of models and the increasing requirements for higherfiltering accuracy in the real world, tranditional filtering methods can’t meet therequirements in the practical applications. Particle filter, as a new nonlinear filteringmethod, is not limited to model characteristics and noise distribution, having ability tosolve nonlinear filtering problems more effectively. Therefore, PF has been widelyapplied in many fields. However, there still exist some drawbacks in the algorithm ofparticle filtering. To improve the performance, the research work on particle filter isvery meaningful both in the sense of theory and the application in practice. This paperaddressed some critical problems in the applications of particle filter under thebackground of target tracking. The research is mainly concentrated in the followingcontents:Under the Bayesian filtering theory, the classic three nonlinear filtering methods,extended kalman filter (EKF), unscented kalman filter (UKF) and particle filter (PF) arestudied based on the representation of dynamic state space model in chapter1. And thenthe advantages, weaknesses and practical limitations of those methods arecomprehensively discussed, respectively. As the core research contents in thisdissertation, the principle of particle filer, design of the proposal distribution, andresampling algorithm are elaborated in detail, then analyzing the particle degeneracyand sample impoverish problem thoroughly in chapter2.Considering how to design better proposal distribution for solving particledegeneracy and sample impoverishment, the design principle and methods based on thestate correction process by nonlinear filtering algorithms are expounded in chapter3.Firstly, RTS optimal smoothing theory are introduced into nonlinear system model,combining with the posterior estimates obtained by iterated EKF filtering, a newproposal distribution function RTS-IEKF was designed. By virtue of new observations,the accuracy and stability of the generated particles are evidently enhanced, thus the proposed algorithm can effectively avoid particle degeneracy and improve theestimation accuracy. Simulations proved that the proposed algorithm has betterperformance in noisy environment.The improvement on resampling is a key issue in particle filter. Based on thecomprehensive research on current resampling techniques, an improved particle filterbased on fine resampling algorithm for general case was proposed in chapter4. Byintroducing distance-comparing and optimized combination strategy, PF-FR canovercome the defects in resampling of generic PF-SIR, making improvement on thestructure of the particle system. As a result, PF-FR can express the posterior PDF of thestate more accurately. Simulative results demonstrated the effectiveness of the PF-FR.In the application of target tracking, by combining ASIFT feature with colordistribution of the target, we construct a new robust measurement model in chapter5.Based on the new hybrid model, a novel target tracking algorithm under the frameworkof particle filter was proposed to realize target tracking more effectively, especially inthe scene existing objects or background with similar color. Video sequences in realworld were used to evaluate the effectiveness and better performance of the proposedalgorithm.Finally in chapter6, the particle fiter tracking method are extended and applied inFourier Telescope System for the requirements of tracking quickly-moving target in thefuture. Based on an idea of designing a set of controller in the transmit system of FTS, anovel approach process was put forward based on the feedback results of particle filtertracking.