Sequential Probabilistic Descriptor via Uncertainty-Aware Multi-Modal Fusion for Safety-Critical Place Recognition

Abstract

The success of loop closure and global map consistency largely relies on robust place recognition, making it a safety-critical capability for mobile robots in autonomous navigation. However, most existing approaches prioritize recognition accuracy while overlooking uncertainty estimation, which may lead to overconfident false predictions and compromise navigation safety. To address this limitation, this letter proposes a sequential multi-modal probabilistic descriptor for place representation, explicitly modeling uncertainty at both the modality and decision levels. Specifically, sequential image and LiDAR inputs leverage temporal information and are encoded as probabilistic embeddings, followed by an uncertainty-aware fusion. By adaptively balancing the contribution of each modality, prediction conflicts caused by sensor degradation are mitigated. Predictive uncertainty is further estimated through probabilistic combination rules, providing reliable confidence measures that allow downstream tasks to filter out high-uncertainty predictions. Extensive evaluations on NCLT, Oxford Robotcar and WildScenes datasets are conducted, covering long-term tasks, seasonal changes, illumination variations, and unstructured natural environments. Results show that the proposed method achieves superior performance across diverse conditions and provides well-calibrated uncertainty estimates that enable reliable failure detection, highlighting its potential for deployment in safety-critical scenes.