Study On Multi-target Tracking Algorithms Based On Random Finite Set

Multiple target tracking (MTT) is an important theoretical and practical problem. The researchers all over the world have been engaged in the area for the recent decades. Many plentiful and substantial results have been achieved, which have been widely applied to military fields such as ballistic missile defense, air reconnaissance and early-warning, battlefield surveillance, etc and some civil fields such as intelligent vehicle system, air traffic control, traffic navigation and robot vision system, etc. However, with the high-technique developing quickly, the tracking environment has changed significantly. The traditional theory and technique of multi-target tracking can not face the new challenges. More and more complex requirements of the theoretic development have been set forth by various applied systems. So the study on multiple target tracking based on the new progress in statistical inference theory is of academic and realistic significance.In recent years, the approaches to MTT based on random finite set (RFS) have attracted more attentions. In these approaches, the collection of individual targets is treated as a set-valued state, and the collection of observations received by sensors is treated as a set-valued observation. So the problem of dynamically estimating the states of multi-target in the presence of clutter can be cast in a Bayesian filtering framework by the RFS formulation. As a result, the process of data association in the traditional approaches to MTT can be avoided. The RFS-based approaches to MTT are studied deeply in this dissertation. The main contributions are as follows:1 Multiple target tracking in the clutter with unknown modelMost of the existing approaches to MTT assumes that the clutter model is known. But in some practical applications, the clutter model might become unknown due to many random disturbances. Therefore, an extended probability hypothesis density (PHD) filter is proposed for the problem of MTT in the clutter with unknown model. In our method, the clutter distribution is first estimated as finite mixture models (FMM) via either expectation maximum (EM) or Markov Chain Monte Carlo (MCMC) algorithm. Then, the estimated model is used directly in the PHD filter to perform multi-target detecting and tracking. The simulation results show that the proposed approach work effectively although the clutter model is unknown. And it outperforms the naive PHD filter which does not estimate the clutter distribution significantly.2 Tracking Partly Resolvable Group TargetsMost of the existing approaches to the problem of group targets tracking (GTT) require the assumption that the individuals comprising of a group are completely resolvable. Under this assumption, the GTT problem can be first converted to the problem of tracking all the individuals involved in the group. But in many practical application, the resolution of the sensor may be not sufficient. It would lead to that the individuals involved in a group is partly resolvable. The existing approaches can not work in such a case. Therefore, we proposes to track the partly resolvable groups using sequential Monte Carlo probability hypothesis density (SMC-PHD) filter. The estimate of the number and the states of the groups, rather than the individuals, is directly derived by the algorithm. The state of a group here consists of its centroid state and shape. In order to estimate the number and the states of the groups, the proposed algorithm fits the distribution of the resampled particles of the SMC-PHD filter by application of Gaussian mixture models (GMM), whose component number and parameters correspond to the number and the states of the groups. EM and MCMC algorithms are respectively used to estimate the parameters of the GMM. The component number of the GMM is derived by the strategy of pruning, merging and splitting. The simulation results show that the proposed approach can track the partly resolvable groups effectively. Besides, the MCMC algorithm outperforms the EM algorithm significantly in extracting the number and the states of the groups.3 Multiple-Model Probability Hypothesis Density (MM-PHD) SmootherFor the problem of tracking multiple maneuvering targets in clutter, the performance of the existing MM-PHD filter is usually unsatisfied due to unknown target number, unknown the association relationship of targets and sensor measurements, unknown motion models, clutter and sensor measurement noise, etc. Therefore, by integrating the MM-PHD filter with the smoothing algorithms, an MM-PHD forward-backward smoother is proposed in this paper for tracking multiple maneuvering targets in clutter. To avoid the use of complex RFS theory, the backward updated equation of the MM-PHD smoother can be derived according to the physical-space explanation of the PHD. Since the MM-PHD forward-backward smoother involves several integral operations, this leads to its recursion analytically intractable in the nonlinear and non-Gaussian conditions. Thus the sequential Monte Carlo (SMC) method is used to implement the smoother. The simulation results show that the proposed MM-PHD smoother significantly outperforms the MM-PHD filter in estimating the number and states of the multiple maneuvering targets. 4 Multiple "non-cooperation" Target Tracking and Multi-sensor Space RegistrationMost of the existing approaches to sensor space registration depend on the measurements of common (or called "cooperation") targets. However, these approaches can not be used directly in the presence of the multiple "non-cooperation" targets (it means that the association relationship of the targets and the measurements is unknown) and clutter. Therefore, a novel algorithm based on PHD filter is proposed for multiple "non-cooperation" targets tracking and multi-sensor space registration in this paper. The number and states of the multi-targets and the measurement biases of all sensors can be simultaneously estimated by this method without the complicated data association. Then the SMC method is used to implement the proposed algorithm. The simulation results show that the proposed method can estimate and compensate the sensor biases relatively accurately. As a result, it significantly outperforms the naive PHD filter, which does not involve the process of the sensor registration, in estimating the multi-target number and states.