Motion-Adaptive Online Loop Closure Detection via Foundation Descriptors and Geometric Consistency

This paper proposes a robust visual loop closure detection pipeline for outdoor environments, utilizing DINOv3-SALAD, a foundation Transformer-based model. Given a sequence of outdoor images, an adaptive temporal buffering mechanism dynamically rejects neighboring frames when feature similarity drops. This effectively prevents false positive loops caused by variable vehicle velocities, which result in inconsistent rates of visual change between consecutive frames. To obtain geometrically valid and locally optimal loop candidates, the visually retrieved pairs are further verified using trajectory-based distance ratios and filtered via 1D Non-Maximum Suppression (NMS). The resulting loop closures are then quantitatively and qualitatively compared with a traditional fixed-interval retrieval approach to evaluate the accuracy, robustness, and effectiveness of the proposed method in handling dynamic outdoor conditions.