A Multimodal Loop-Closure Detection Framework with Dynamic-Object Suppression

In autonomous driving scenarios, the presence of dynamic objects can significantly degrade the precision and recall of loop-closure detection. To further improve robustness, this paper proposes a loop-closure detection framework with dynamic-object suppression. The framework jointly exploits camera and LiDAR sensors to obtain rich environmental information. More importantly, we introduce a probabilistic dynamic suppression mechanism at the local-descriptor level. Based on cross-path consistency between the image and bird’s eye view (BEV) branches and the temporal persistence of BEV grid-cell activations, we construct two complementary pieces of evidence for detecting anomalous, potentially dynamic content. These two cues are then fused in probability space using a Bayesian formulation to produce interpretable dynamic posterior probabilities for each local feature. The resulting probabilities are converted into soft suppression weights that reweight the local descriptors before NetVLAD aggregation, yielding cleaner local descriptors for loop-closure retrieval. Compared with the baseline, the proposed method improves not only average precision but also Recall@100 and F1-score. On the public KITTI-360 dataset, our approach achieves significant performance gains. In particular, on the challenging KITTI-360-00 sequence, average precision (AP) increases from 0.895 to 0.945, and similar improvements are observed on other sequences, demonstrating the effectiveness of the proposed framework.