UC San Diego

Modern estimation theory for engineering applications is increasingly characterized by multimodal, high-dimensional state distributions and nonlinear, high-accuracy measurements with complex data associations. In this dissertation, we address the challenges of sampling-based multi-object tracking (MOT) and localization in complex 3D environments. I first focus on enhancing sampling efficiency in passive MOT scenarios. To this end, I incorporate particle flow (PFl) techniques into a belief propagation (BP) framework for sequential estimation of an unknown number of states in the presence of measurement-origin uncertainty. In particular, I develop a "flow-induced'' proposal PDF for importance sampling that consists of a weighted mixture of parallel flows performed for multiple candidate measurements. Particle flows are either modeled using deterministic or stochastic spatio-temporal processes. The resulting method for MOT is evaluated with synthetic data and in a challenging passive acoustic undersea tracking problem where the 6-D state of multiple whales is tracked based on 1-D time-difference of arrival (TDOA) measurements provided by pairs of hydrophones on a volumetric array. This thesis also introduces an active planning method for cooperative localization. Here, sensor positions are dynamically controlled to maximize the information gain of future measurements. The control policy is derived by minimizing the trace of approximate inverse Bayesian Fisher information matrices (FIMs). Numerical results demonstrate that particle flow significantly improves sampling efficiency and estimation accuracy, while active planning leads to enhanced localization performance.