Stochastic methods for bayesian filtering and their applications to multicamera multitarget tracking

Abstract

Target tracking is an important key technology for many military and commercial applications. The tracking problems are usually formulated by using the state space approach for discrete-time dynamic systems. Under this framework, the tracking is to estimate the state x_t of target at time t, given the measurement sequence y_{1:t} of sensor from time 1 to t, or equivalently to construct the conditional probability density function p(x_t|y_{1:t}). The theoretical optimal solution is provided by the recursive Bayesian filter. However, for multi-sensor multi-target tracking, there are many challenges to extend the single-sensor single-target Bayesian filter. In this thesis, the focus is on extending the Bayesian filter to multi-camera or multi-target visual tracking.First, a spatio-temporal recursive Bayesian filter is formulated for tracking a target using multiple cameras. We propose an adaptive mixed particle filter for the implementation of the spatio-temporal recursive Bayesian filter for the dynamic system. In particular, the mixed importance sampling strategy is used to fuse temporal information of dynamic systems and spatial information from multiple cameras. It is adaptive in sense that it automatically ranks data from multiple cameras and assigns weights according to data's quality in the fusion process. The results show that this method is able to recover a target's position even when it is completely occluded in a particular camera for some time.Second, a multi-target Bayesian filter, the probability hypothesis density (PHD) filter, is designed to track unknown and variable number of targets in image sequences. Because the dimensions of state and observation are time-varying during the tracking process, the PHD filter employs the random finite set representation of multiple states and multiple measurements and the PHD is the 1st order moment of random finite set. The PHD filter is implemented using two methods: both particle filter and Gaussian mixture. For the particle PHD filter, two importance functions and correspondent weight functions are proposed for survival targets and new-birth targets, respectively. It is shown in the thesis that the importance function for survival targets theoretically extends the optimal importance function of the linear Gaussian model from single-measurement case to measurement-set (multi-measurement) case. Whereas the importance function for new-birth targets is a data-driven method which uses the current measurements in the sampling process of the particle PHD filter. For the Gaussian mixture PHD filter, a scene-driven method which incorporates the prior knowledge of scene into the PHD filter is presented. The results show that these PHD filters are able to track a variable number of targets and derive their positions in image sequences.This work suggests that stochastic methods for Bayesian filtering are powerful means for multi-sensor multi-target tracking.