摘 要:介紹了粒子濾波基本原理,針對粒子濾波計算量大和難以用硬件實現等缺點對粒子濾波算法進行了改進,使其平均計算周期縮短為原來的90%,應用DSP實現了粒子濾波算法。改進粒子濾波算法主要優化了原粒子濾波算法中權值計算、重采樣和輸出步驟,使其計算速度和濾波精度有所提高。這種改進粒子濾波算法在DSP系統中進行仿真,結果證明它具有速度快,精度高的優點。關鍵詞:粒子濾波; 硬件實現; 權值計算; 數字信號處理器
中圖分類號:TN911-34文獻標識碼:A
文章編號:1004-373X(2010)18-0009-04
Improved Particle Filter Algorithm Based on DSP
FENG Jun-xiang, ZHANG Jian, BO Chao
(School of Electronics and Information, Jiangsu University of Science and Technology, Zhenjiang 212003, China)
Abstract: Particle filter is based on the Monte Carlo and recursive Bayesian estimation which has special advantages in dealing with the nonlinear and non-Gaussian problems. However, both the enormous computations and low speed restrict its implementation in real-time system. At first, the basic theory of particle filter is introduced. Then, the particle filter algorithm is improved for solving the disadvantages of enormous computations and hard implementation of hardware, which reduced the average cycle to 90%. At last, particle filter algorithm is fulfilled through DSP. Compared with original particle filter algorithm, the improved algorithm optimizes the steps of computation of weights, resample and output,which improve the calculator speed and filter precision. The improved particle filter algorithm testifies its advantages of fast speed and high accuracy through carrying out simulation in DSP system.Keywords: particle filter; hardwareimplementation; computation of weights; DSP
0 引 言
粒子濾波(particle filtering,PF)[1-2]或Monte Carlo粒子濾波(MCPF)是以重要性采樣(importance sampling,IS)和序貫重要性采樣(sequential IS,SIS)為基礎的序貫Monte Carlo(sequential MC,SMC)方法,因此又稱為SMC濾波,1999年正式提出PF稱謂[3],該名稱現已廣泛采用。由于PF算法在理論上對高維非線性、非高斯動態系統的狀態遞推估計或概率推理等問題都不具敏感性,因此,在復雜問題的求解上它表現出突出的優勢。但是到目前為止,在實時信號處理領域,粒子濾波算法幾乎沒有得到實際應用,這主要是因為粒子濾波算法本身較復雜,運算量大,需要存儲的空間大。某些改進粒子濾波算法雖然在一定程度上提高了粒子濾波算法的精度,卻使得粒子濾波算法更加復雜,實時性很差。
在此,首先介紹了標準粒子濾波算法,之后從硬件實現的角度出發,將粒子濾波權值計算中的權值歸一化部分合并到重采樣計算和輸出計算步驟中,并且改進了權值計算方法,以非線性非高斯系統為例,驗證了改進算法的精度,結果說明,改進算法更適合硬件實現,一定程度上在提高了算法運算速度的同時,提高了算法的濾波精度?!?br>