The Control of Mobile Sensing Platforms to Perform Estimation of Mobile Targets

The motivation for this thesis is due to the proliferation of uninhabited mobile sensing platforms capable of performing information-gathering tasks autonomously. These platforms are typically equipped with onboard computing, inter-platform communication, and target sensing capabilities. We are interested in leveraging these capabilities to perform state estimation on mobile targets. The primary challenge is to design a control policy that would steer the observers to best estimate the targets' states. This thesis is devoted to the problem of designing control policies for state estimation, where estimation is not a means but an end to the control problem. We first address the scenario of a single observer tracking a single target, followed by the problem of multiple observers tracking multiple targets. The formulation of this thesis differs from that of the previous works in many respects. First, we are interested in modeling a sensor that would provide a noisy measurement of the target's location, given that a measurement is made. The noise would increase as the distance between the target and the observer increases, and there would always be a non-zero probability for a missed detection. This models the behavior of an electro-optical sensor running a computer vision algorithm for target detection and measurements. Second, we presume a highly mobile target being observed by an observer with dynamic constraints. This typifies the relative dynamics between the target and the sensing agent of interest in today's applications. Third, we are interested in maintaining "good" steady state tracking performance. This differs from that of the previous works where a finite horizon problem or a search scenario is considered. To this end, we pose and solve two problems: a binary and a one-dimensional search-and-track problem. While both problems are mostly academic, they both capture the key features of interest in the target tracking problem that were not addressed by the previous efforts. To solve the problems we employ the approach of dynamic programming and the technique of value iteration. For the binary tracker, a direct application of value iteration was sufficient. However, for the more complicated one-dimensional formulation, we use the partially observable Markov decision process (POMDP) as a framework and the technique of point-based value iteration (PBVI) as an algorithm to find the solution. The results show that computation complexity is effectively circumvented while the main features of the problem are maintained. Success from these formulations offers promise for success in a real-world implementation. The problem of multiple observers tracking multiple targets delves into the issue of distributed decision-making. At the heart of every problem on distributed decision-making is the issue of information patterns, and this is also the case for the problem of distributed target tracking. Nearly all previous formulations of this problem presume a perfect communication channel. Under this condition, work has been devoted to understand how to distribute the dynamic programming problem across multiple agents while using minimal communication efforts. The formulation in this thesis assumes a communication constraint at the inception. It is well known that the communication constraint is a notoriously difficult concept to quantify, and so the approach of model predictive control (MPC) is chosen for its ability to handle constraints. Since the contribution would be specific to the formulation, we model the system to closely resemble that of the fleet of unmanned aerial vehicles (UAVs) that is in use by the Center for Collaborative...