Research And Implementation Of Particle Filter Algorithm

In recent years, based on Monte Carlo, particle filter is a nonlinear and non-Gauss filtering algorithm, which completely break through the traditional Kalman theory framework, and can be applicable to any state space model of nonlinear system. Besides, the precision of particle filter can approximate the optiomal estimation. Particle filter is a cross discipline between statistical and simulation theoretical disciplines and model signal processing, which is very important to theoretical and practical value. Particle filter have a wide application in radar target tracking, speech signal enhancement, sensor fault diagnosis, the inverted swing control system, statellite navigation and economics and biological control.In this thesis, the development background and significance of the particle filter is summarized, starting from the equation of state, throuth the Bayesian theory and Monte Carlo sampling method of particle filter. According to the standard particle filter, there are some defect, such as poor estimation accuracy, particle degeneration, the important function of particle hard to choose, large amount of calculation and bad real-time. Combined with Extended Kalman particle filter and Unscented Kalman particle filter, this thesis proposed an important function by Gauss-Hermite Kalman filter to realize particle filter. The algorithm complexity of partical filter is lower than the Unscented Kalman particle filter, and its performance is better than the Extended Kalman particle filter, which provides a choice as the importance function of Partical filter.This thesis study on the four basic resampling algorithm and compare the four kinds of algorithm by Matlab simulation, and improve the defect of not completely abandon the smaller paritcles. According to the adaptive resampling algorithm, this thesis improved resampling algorithm for linear resampling. This algorithm is the loss of precision, but is very efficient in hardware implementation. Combined the method of the optimization of weight and adaptive resampling algorithm, this thesis proposes a new algorithm to avoid a lot of data sorting problem. So it can reduce the complexity of the algorithm, and its performance is close to the original weights optimization algorithm.Finally, this thesis research on circuit design of partical filter, finding the circuit have large amount of calculation and bad real-time problems. The method is proposed to avoid the weight normalized to reduce division operation, and to solve the problem of poor real-time by the parallel circuit design. In the circuit design, the particle circuit is effectively divided into sampling circuit, weight calculation circuit and resampling circuit. The main error of particle filter circuit design exists in exponent arithmetic, and bad real-time problems mainly exist in the resampling algorithm. The circuit design provides the basis for further improve the performance of particle filter hardware design.