Conditional Posterior Cramer-Rao Lower Bound and Distributed Target Tracking in Sensor Networks

Sequential Bayesian estimation is the process of recursively estimating the state of a dynamical system observed in the presence of noise. Posterior Cramer-Rao lower bound (PCRLB) sets a performance limit on any Bayesian estimator for the given dynamical system. The PCRLB does not fully utilize the existing measurement information to give an indication of the mean squared error (MSE) of the estimator in the future. In many practical applications, we are more concerned with the value of the bound in the future than in the past. PCRLB is an offline bound, because it averages out the very useful measurement information, which makes it an off-line bound determined only by the system dynamical model, system measurement model and the prior knowledge of the system state at the initial time. This dissertation studies the sequential Bayesian estimation problem and then introduces the notation of conditional PCRLB, which utilizes the existing measurement information up to the current time, and sets the limit on the MSE of any Bayesian estimators at the next time step. This work has two emphases: firstly, we give the mathematically rigorous formulation of the conditional PCRLB as well as the approximate recursive version of conditional PCRLB for nonlinear, possibly non-Gaussian dynamical systems. Secondly, we apply particle filter techniques to compute the numerical values of the conditional PCRLB approximately, which overcomes the integration problems introduced by nonlinear/non-Gaussian systems. Further, we explore several possible applications of the proposed bound to find algorithms that provide improved performance. The primary problem of interest is the sensor selection problem for target tracking in sensor networks. Comparisons are also made between the performance of sensor selection algorithm based on the proposed bound and the existing approaches, such as information driven, nearest neighbor, and PCRLB with renewal strategy, to demonstrate the superior performances of the proposed approach. This dissertation also presents a bandwidth-efficient algorithm for tracking a target in sensor networks using distributed particle filters. This algorithm distributes the computation burden for target tracking over the sensor nodes. Each sensor node transmits a compressed local tracking result to the fusion center by a modified expectationmaximization (EM) algorithm to save the communication bandwidth. The fusion center incorporates the compressed tracking results to give the estimate of the target state. Finally, the target tracking problem in heterogeneous sensor networks is investigated extensively. Extended Kalman Filter and particle filter techniques are implemented and compared for tracking a maneuvering target with the Interacting Multiple Model (IMM). Conditional Posterior Cramer-Rao Lower Bound and Distributed Target Tracking in Sensor Networks by