Characterization and optimization of the accuracy of mobile robot localization

摘要

In this work, we describe methods for analytically characterizing and optimizing the accuracy of robot localization. The first problem we address is that of providing performance guarantees for localization. We focus on three classes of localization algorithms: Cooperative Localization (CL), Simultaneous Localization and Mapping (SLAM), and Cooperative SLAM (C-SLAM). In each of these cases, we analytically derive upper bounds on the localization uncertainty, as functions of the relevant system parameters. The resulting closed-form expressions allow us to study the asymptotic properties of localization and examine the impact of several factors on the localization accuracy. The second problem we address is optimal resource utilization during CL in robot formations. We propose a methodology for selecting the optimal set of measurements to process at each time-step, to attain the highest possible localization accuracy given the robots' limited computational and communication resources. Our approach is based on first expressing the localization accuracy as a function of the rate at which each of the available sensors is utilized, and then formulating a convex optimization problem, to determine the optimal rate for each of the sensors. The convex nature of the optimization problem (semi-definite programming) allows computing a globally optimal solution using efficient minimization routines. Finally, the third problem we focus on is that of optimal resource allocation during vision-aided inertial navigation. One key challenge in this domain is the very large number of feature measurements, which can overwhelm the limited computational capabilities of a real-time system. Our approach to this problem is based on the observation that the vast majority of features can only be tracked in a small number of frames (transient features). We propose an algorithm that can optimally process the measurements of such features, at computational complexity only linear in their number. We then show how this algorithm can be incorporated into a two-layer localization system that utilizes transient features, as well as the loop-closure information that becomes available when a robot re-visits an area. The merging of these two types of positioning information produces pose estimates that are available in real time and have bounded longterm errors.