Research On The Particle Filter Tracking Methods

The most achievements of the signal processing theory are still on the random signal with Gaussian assumption and the using the linear model. However the statistical characteristics of the signal are not always accords with Gaussian hypothesis, and the complex signal cannot be correctly described by linear model. Then the non-Gaussian and nonlinear signal processing have been the focus of researchers.Based on state space model, the recurred Bayesian estimation theory is the theoretical framework of Kalman filter and Particle Filter. Kalman filter can only be used for the linear Gaussian system. The extended Kalman filter, unscented Kalman filtering algorithms all have the drawbacks that the computation complexity will increased rapidly with higher dimensions of states , and the only the 1 and 2 order statistics information are used. In the severe nonlinear and non-Gaussian circumstances, the filter performance fell sharply. The Particle Filter uses the Monte Carlo method to implement the recursive Bayesian filter. The key idea is to use weighted random samples (particles) to approximate the posteriori probability density. This approximate method is applicable for various nonlinear non-Gaussian systems, and can effectively overcome the defects of the previous filtering algorithm. In this paper, a detailed analysis of the principle of the particle filter is presented, and some new techniques are derived to improve the performance of particle filter, especially in the particle filter tracking problem in positioning and tracking application. The main work and innovation summary are as follows:Analysis of the principle and resampling algorithms and the comparison of several resampling algorithm are presented. A new resampling algorithm based on artificial immune algorithm is derived, which can overcome sample impoverishment effectively. The paper introduces the polynomial resampling algorithm and the other algorithm based on it, including the residual resampling, stratified resampling, the systematic resampling. With the analysis of the principles and resampling, this paper discusses the standard for a good resampling algorithm, and analyzes the performance of these resampling algorithms. An example of the maneuvering target tracking problem is given to evaluate the performance and the computational complexity, proving the validity of the theoretical analysis. To overcome the sample impoverishment, a new resample algorithm is proposed based on artificial immune algorithm. The new resampling algorithm can increase the diversity of the particles after the resample process, and can overcome the sample impoverishment effect.The particle filter algorithm provides posteriori probability density, including all the statistical information of state. Then a new Bayesian estimation based on fourth-order risk function is developed. The large quantities'particle filters, with corresponding weights of random sample, can be used to approximate the system state posteriori probability density function, and all the statistical information of the system state can be derived from it. But most of filtering algorithms use MMSE estimation, only using the one and two order statistical information. In the severe nonlinear non-Gaussian circumstances, the MMSE estimation performance will be serious decline. In order to make full use of statistical information, this paper defined a new kind of four orders risk function, and a new Bayesian estimation is give based on the four-order risk function. Although the new estimation requires some integration between the parameters estimation and the posteriori probability distribution, it can be easily implemented with the particle filter. Supposing the distribution of noises is the Alpha stable distributions, a univariate non-stationary growth model is given to test the performance using particle filter, which uses the new Bayesian estimation based on the four-order risk functions. Simulations show that the new algorithm outperforms the standard particle filter based on the MMSE.Position tracking problem is an important kind of applications in which the Particle Filter can be applied. The performance of many key techniques can be improved by using particle filter.In TDOA localization technology the precision clock synchronization is very important. A new TIE online estimation algorithm based on the GPS and particle filter are developed. The location accuracy of TDOA localization system heavily depends on time measurement precision, so the precision clock becomes the one of the important factors affecting positioning accuracy. Using GPS clock real-time estimation crystals of clock time intervals and compensate error can achieve a more accurate clock. Research shows that crystals senescence rate obey logarithm aging model, using the linear model describes the previous studies of crystals were not very accurate drift. Based on the aging model established new logarithm crystal oscillator model, and the clock drift particle filter for the nonlinear model. Compared with the existing methods, based on the particle filter algorithm can obtain better TIE the estimated performance.For indoor wireless location technology, the nonlinear tracking algorithm is proposed based on the particle filter and the probability density distribution. The RSS based wireless location has drawn lots of attention because of its low cost, convenience for deployment. Histogram method the kernel function method are presented for approximate the probability distribution, as the fingerprints of certain location. These methods make full use of the information RSS samples, obviously improving the positioning accuracy of tracking. The RSS measurements are directly used as the measurment of the track model. The particle filter is applied on the nonliear model and the simulation shows that the algorithm outperforms the old ones, using the log normal model or the RSS based fingerprints.For wireless sensor networks of tracking problem, this paper proposes a TOA/RSS or TDOA/RSS hybrid positioning tracking algorithm. The TOA measurements require a strong node and result in the good performance, while the RSS is easily got with bad perfomance. This paper presents a TOA/RSS hybrid tracking algorithm to make full use of the advantages of the two location technique. The two measurments are all used as a system measurements and a model is constructing. Simulation shows that the new algorithm can improve the tracking performance and can only use one TOA measurements to work.