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

In visual navigation, Simultaneous Localization and Mapping (SLAM) faces the pivotal challenge of loop closure detection, which is vital for refining position estimates and map construction. Current approaches relying on deep learning-based global descriptors struggle with robustness and interpretability. To overcome these limitations, we propose a novel method that integrates partial semantic segmentation with traditional location recognition networks through a weighted fusion mechanism. By harnessing the synergy of semantic and spatial information, our approach provides deeper insights into image content and spatial relationships. The carefully crafted weighting scheme enables a more comprehensive assessment of image similarity, considering both the “what” and “where” of image features. Experimental evaluations conducted on the Pittsburgh 250k dataset, comprising an extensive collection of 250,000 images, consistently showcase the effectiveness of our fusion strategy. Across all three tested backbone networks, we observe a notable improvement of over 3 % in recall rate for loop closure detection. Notably, when employing MobileNet as the backbone, the enhancement is even more pronounced, surpassing 5% with an optimal configuration featuring a semantic vector weight of 0.94 and a location network weight of 0.06. This significant achievement not only underscores the robustness and accuracy gains achievable through our approach within SLAM systems but also highlights its potential as a versatile strategy for semantic-spatial integration, with promising applications in various computer vision tasks that require advanced spatial-semantic comprehension.