Research On Sequential Monte Carlo Methods For Nonlinear Filtering Techniques

In reality, stochastic dynamic systems are mostly nonlinear and non-Gaussian.So the nonlinear filtering problem is very popular and many applications are concerned with it. Target tracking is one important field among its applications. Target tracking is the estimation of unknown target kinematics state based on uncertain measurements from sensors. Recently, particle filter, also known as Sequential Monte Carlo Methods, has become popular tools to solve tracking problems. The popularity stems from their simplicity, flexibility, and ease of implementation, especially their powerful ability to deal with general nonlinear and non-Gaussian estimation problems. This nonlinear filtering technique is a numerical method based on simulation and to use discrete hidden Markov chain for modeling, the system model describes the evolvement of the unknown target state over time and the measurement model associates the available measurements to the target state. Using the past and present measurements, based on prior information, the prediction and update step are performed, an approximate distribution for the target state is attained.In this dissertation, Sequential Monte Carlo methods and the principle of target tracking under the Bayesian framework are studied. The methods for improving the particle filter are discussed in detail. Also the convergence properties of particle filter are analysed. The simulation experiments are conducted to verify, for linear and Gaussian system, the exact estimation methods have better results than particle filter. However, for nonlinear and non-Gaussian dynamic system, particle filter has obvious dominance and it could improve the effect of filtering evidently. Also we demonstrate Rao-Blackwellized particle filter has good results for voice enhancement.For the multi-sensor target tracking problems, the data fusion techniques are discussed and the Cross-Sensor and Cross-Modality (CSCM) data fusion algorithm is presented, which can deal with multiple sensors with different types and modalities. And particle filter is applied for nonlinear estimation to track a moving mobile robot. In order to find the state which describes the target position accurately, 3 different methods are used for estimation: the best particle, the weighted mean and the robust mean. Also 3 basic resampling schemes are compared. The experimental results show the feasibility and the effectiveness of the data fusion algorithm.On visual tracking, the visual tracking techniques, including camera systems, color distribution and the other correlated knowledge are discussed. And a particle filter algorithm is presented based on color histogram to track a moving target which can deal with rotation, scale changes, variations in the light source and partial occlusions. So it can track the target with robustness. The proposed method is based on particle filter, integrated with color histogram in the measurement model, and the system model is a second-order autoregressive process. The tracked target can be rigid or non-rigid. Also the method can run in real-time.For the multi-target tracking problems with complex background, the data association techniques are discussed, we combine the target detection algorithm with particle filter, use color histogram as observation model, and the global nearest neighbor (GNN) performs data association. A multi-target tracking algorithm is presented based on particle filter. The proposed algorithm is robust to the problems of the appearance and disappearance of targets, the similar appearance of targets, the cross movement of targets and short-time occlusion.For the performance evaluation of target tracking based on particle filter, a quantity approach is developed which can evaluate the quality of tracking algorithm. The strategy is theoretically based on precision and recall, complementary to the logical assessment of a visual tracking system's performance. Via designing the experiments, the optimal parameters are chosen. All experiments have solely used single-target tracking. However, the change from single-target tracking to multi-target tracking is also evident.