Research On Particle Filter And Its Application To Target Tracking

With the rapid development of control technology and computer technology, nonlinear filtering technology has been found wide applications in signal processing, automatic control, computer vision, wireless communication, aerospace and target tracking and recognition areas. Based on the framework of Bayesian theory, particle filter (PF) has become the focus of optimal estimation for nonlinear and non-Gaussian dynamic system. With the background of the national natural science foundation research, the dissertation mainly investigates particle filter and its application to target tracking.1. For the particle degeneracy problem with standard particle filter, an improved PF is proposed. A modified iterated extended Kalman filter is used to generate importance density function and the Markov chain Monte Carlo step is adopted to maintain the diversity of particles. Based on the glint noise statistical model, the improved particle filter is verified to be effective by the computer simulations. And then, for maneuvering target tracking, a novel interactive multiple model PF is proposed, which uses an improved PF in multiple model. The performance of the algorithm is illustrated via simulations under different conditions. Finally, for the nonlinearity and unobservability problems with bearings-only tracking, a new distributed multi-sensor optimal information fusion PF is proposed. As far as performance analysis is concerned, a theoretical Cramer-Rao lower bound (CRLB) is derived. The computer simulation results indicate that improvements of both fault tolerance and robustness properties are achieved with the proposed algorithm.2. For the fixed number target tracking problem, works are concentrated on multiple targets tracking problems based on data association and particle filter. Firstly, a novel Gaussian particle joint probabilistic data association filtering algorithm is proposed, which introduces Gaussian particle filtering concept to the joint probabilistic data association (JPDA) framework. The algorithm subsititutes Gaussian particles for Gaussians in the JPDA filter. Computer simulation results show that the tracking precision of the proposed algorithm is improved and that the filtering variance is reduced. Secondly, as far as uncertainty of the target dynamic model, a novel algorithm based on fuzzy clustering and PF is proposed for multi-target tracking. The joint association probability matrix is constructed by using the maximum entropy fuzzy clustering method. Then, particle filter is employed to update each target state independently. The effectiveness of the algorithm is verified by the computer simulations.3. For the unknown and time-varying number target tracking problem, multiple targets tracking theory based on Finite Set Statistics (FISST) is studied. A new Gaussian mixture particle PHD filtering algorithm is presented. The PHD is approximated by a weighted mixture of Gaussians as is the case in the Gaussian mixture PHD (GM-PHD) filter. The Quasi-Monte Carlo integration and Gaussian particle filter (GPF) are applied for predicting and updating the Gaussian components. The resulting filter does not require clustering techniques to determine the target states and the Gaussian particle approximation is asymptotically consistent. Computer simulation results show that the tracking performance of the proposed algorithm is improved.4. For the problem that PHD filter estimates the target number with a corresponding high variance, the cardinalized PHD (CPHD) filtering algorithms are discussed. Firstly, the Gaussian mixture particle PHD filtering concept is extended to the CPHD filter. A novel Gaussian mixture particle CPHD filter is proposed, which uses a set of Gaussian particle filters to jointly estimate the posterior PHD and the posterior distribution of the number of targets. In addition, for mixed linear/nonlinear state space models, a new particle CPHD filtering algorithm is proposed, which combines particle filter and Kalman filter to predict and estimate the states of multiple targets to ehhance the estimating performance of the PHD and cardinality distribution. The target state estimates are extracted by utilizing the kernel density estimation theory and mean-shift method. Simulation results are presented to demonstrate the improved performance of the proposed filtering algorithms.