Passive map learning and visual place recognition

摘要

An autonomous mobile robot needs to learn about its environment while pursuing other tasks. This dissertation investigates the problem of passive map learning, where the robot is not allowed to rely on exploration for learning. Forbidding exploration makes the problem more difficult, since the true state of the world cannot always be inferred from the incomplete and noisy information available at any one point in time. If the robot cannot actively explore in order to verify the state of the world, the mapping system will inevitably make errors. Two main ideas contribute to the operation of the passive mapping system described here: (a) robot localization and map updating follows a principle of least commitment, with a proper treatment of estimation ambiguity, and (b) errors in mapping are allowed, but explicitly diagnosed and eventually corrected. The map representation used is path-based, but incorporates metric information to provide the constraints necessary for passive mapping. These metric constraints allow the system to diagnose mapping errors and to correct them. The system has been shown empirically to converge on correct maps of its environment in a realistic simulation. Another important part of map learning is place recognition, or deciding where the robot is in its map. This dissertation also presents a novel image-based method for using vision to aid in place recognition, using image signatures. An image signature is an array of values derived by evaluating a measurement function over large blocks of pixels. Measurements are chosen to be characteristic of a location yet invariant over different viewing conditions. Signature matching is performed quickly by element-wise comparison. Experiments on a large image corpus demonstrate that image signatures give an accurate and efficient method for visual place recognition.