Deep SIMBAD: Active Landmark-based Self-localization Using Similarity-based Scene Descriptor

Abstract

Landmark-based robot self-localization has attracted recent research interest as an efficient maintenance-free approach to visual place recognition (VPR) across domains (e.g., times of the day, weathers, seasons). However, landmark-based self-localization can be an ill-posed problem for a passive observer (e.g., manual robot control), as many viewpoints may not provide effective landmark view. Here, we consider active self-localization task by an active observer, and present a novel reinforcement-learning (RL) -based next-best-view (NBV) planner. Our contributions are summarized as follows. (1) SIMBAD-based VPR: We present a landmark ranking -based compact scene descriptor by introducing a deep-learning extension of similarity-based pattern recognition (SIMBAD). (2) VPR-to-NBV knowledge transfer: We tackle the challenge of RL under uncertainty (i.e., active self-localization) by transferring the VPR's state recognition ability to NBV. (3) NNQL-based NBV: We view the available VPR as the experience database by adapting a nearest-neighbor -based approximation of Q-learning (NNQL). The result is an extremely compact data structure that compresses both the VPR and NBV modules into a single incremental inverted index. Experiments using public NCLT dataset validate the effectiveness of the proposed approach.