Research On Multi-target Tracking Based On Random Finite Set

The multi-target tracking is to estimate states of moving targets by signal processing algorithm. As an attractive field, multi-target tracking can be applied in video surveillance, intelligent transportation systems, radar, robot, etc. Traditional multi-target tracking approaches, for example nearest neighbor and joint probabilistic data association, design single target random filters, choose corresponding observation for every filters, and receive good results. These approaches have been applied in multi-target tracking. However, they often demand many prior conditions or the targets’number should not be changed during tracking. Furthermore, their computational complexity is high because of using data association algorithm.In recent years, people are looking for the multi-target tracking approaches based on random finite set. The most representative algorithm is probability hypothesis density filter proposed by Mahler. In this algorithm, measurements and multiple targets’ states and number are presented by random finite set, and the estimates of states are given by Bayes filter. Compared with traditional multi-target tracking approaches, the algorithms based on random finite set avoid using data association technique, and their computational complexity are controlled efficiently. Furthermore, the targets’ number could be changed during tracking.This dissertation starts with a research of multi-target tracking filter based on random finite set. And it proposes some improvement of probability hypothesis density filter in linear and nonlinear Gaussian system to improve the robustness and estimation precision of the filter. At the same time, multi-target filter based on random finite set is used to tracking multiple speakers, and the filtering accuracy of speakers’states is enhanced.The main work of this dissertation is as follows:(1) This dissertation proposes an improved Gaussian components’ merging algorithm in Gaussian mixture probability hypothesis density filter to enhance the filtering accuracy of the filter while tracking multiple close proximity targets. The Gaussian components received from recursion are merged selectively. Meanwhile, the weights of Gaussian components are used to decide whether the components can be utilized to extract states, and the means and covariances are used to determine whether the components will be merged. Consequently, the proposed algorithm avoids that the components which are used for extracting states are merged. Simulation results show that the new algorithm can enhance the precision and robustness of estimation for multiple target states when the targets move closely. (2) Aiming at nonlinear system model in multi-target tracking, a central difference Kalman-probability hypothesis density filter is proposed to track multiple targets. Multi-target tracking is fulfilled by deriving polynomial approximations with Stirling interpolation formulas, estimating first-order statistical moment of posterior multi-target states with central difference Kalman filter and Gaussian mixture probability hypothesis density filter, and extracting states of targets from the recursion of probability hypothesis density. The advantages of proposed filter are that Jacobian matrix solving is unnecessary and second-order Taylor expansion accuracy can be ensured. Simulation results show that the effect on the algorithm by non-linear system model is reduced, and estimating accuracy of target’s number and states is improved.(3) In connection with nonlinear system model in multiple speakers tracking, a cubature Kalman-probability hypothesis density filter for multiple speakers tracking is proposed. Time difference of arrival for microphone array is taken as measurements, third-degree spherical-radial rule is utilized to compute the multidimensional integral in Bayesian filter of nonlinear system in proposed method, cubature Kalman filter and probability hypothesis density filter is applied to estimate first-order statistical moment of posterior multiple speakers states, and finally multiple speakers tracking of nonlinear Gaussian system is realized while the speakers’states are extracted by recursive updating. Calculating Jacobian matrix of nonlinear system function is no longer necessary in proposed filter and its computational complexity also goes down. Simulation experiments have been taken to analyze the performance of proposed method when detection probability, false speakers’ number, sampling period, speech-signal-to-noise ratio and reverberation time varies. Simulation results show that the proposed method enhances the robustness of tracking algorithm, and improves estimation accuracy of multiple speakers’number and states.