Abstract: In statement tracking, drastic change increases the process noise and accordingly increases the difficulty of self-adaptive filter tracking. Traditional particle filter algorithm has a disadvantage that if change is too drastic, it can not correct errors effectively, which makes the estimation errors cumulate and the tracking system become divergent. Unscented particle filter (UPF) algorithm, which uses an unscented Kalman filter (UKF) for proposal distribution generation within a particle filter framework, can decrease the posterior probability distribution estimation error, enhance tracking effect, but it also increase the computation time. An improved particle filter al-gorithm(PF-RBF) based on radial basis function network (RBFN) is proposed, which aims at improving the sampling process of new particles and reducing the computation time. The algorithm uses RBFN to construct the process model dy-namically from the observations and update the state of the system, which can reduce prior probability distribution estima-tion error and remove the cumulated effect of errors. Compared with UPF, PF-RBF can reduce computation time because it doesn’t contain UKF process. The target tracking experiment results verify that PF-RBF performs better than UKF, PF and UPF whether the observation model is nonlinear or linear. Fur-thermore, the intrinsic property of PF-RBF determines that the change rate of execution time of PF-RBF is less than UPF, so PF-RBF is more suitable for large-scale applications.

Key words: Particle filter, Radial basis function neural network, State tracking