Semantic Information-enhanced Loop Closure Detection for Simultaneous Localization and Mapping

Yushan Huang

Abstract: Abstract——In visual navigation, Simultaneous Localization and Mapping (SLAM) faces the pivotal challenge of loop closure detection, which is vital for refining position estimates and mapconstruction. Current approaches relying on deep learning-based global descriptors struggle with robustness andinterpretability. To overcome these limitations, we propose anovel method that integrates partial semantic segmentation withtraditional location recognition networks through a weightedfusion mechanism. By harnessing the synergy of semantic andspatial information, our approach provides deeper insights intoimage content and spatial relationships. The carefully craftedweighting scheme enables a more comprehensive assessment ofimage similarity, considering both the "what" and "where" ofimage features. Experimental evaluations conducted on thePittsburgh 250k dataset, comprising an extensive collection of250,000 images, consistently showcase the effectiveness of ourfusion strategy. Across all three tested backbone networks, weobserve a notable improvement of over 3% in recall rate for loopclosure detection. Notably, when employing MobileNet as thebackbone, the enhancement is even more pronounced, surpassing 5% with an optimal configuration featuring a semantic vectorweight of 0.94 and a location network weight of 0.06. Thissignificant achievement not only underscores the robustness andaccuracy gains achievable through our approach within SLAMsystems but also highlights its potential as a versatile strategy forsemantic-spatial integration, with promising applications invarious computer vision tasks that require advanced spatial-semantic comprehension.

Submission Number: 169