Localization for Lost Robots through Scene Recognition ( May 2008 )

Mobile robot localization is a common problem in the robotics community, yet many proposed solutions are hardware dependent, requiring various sensors with prior environment information, or solutions that require continuous feature tracking such as the simultaneous localization and mapping (SLAM) framework, where prior motion dynamics or accurate odometry is required. However, these localization techniques are not robust to handle localization for single-camera lost robots, where the system can encounter temporary lack of features to track, thus hindering the system lost without any method to recover and regain correspondence to the previously built map. In this paper, we present a technique that addresses these issues and achieves robust performance in localizing a lost robot through finding correspondence between the current scene and a previously built map.