SeqMamba-MPR: A Spatial-Temporal Mamba Network for Place Recognition Using Sequential Multi-Modal Data

Place recognition is a vital technique in the field of autonomous vehicle localization. Existing LiDAR-Visual fusion methods suffer from excessive computational burden, and their real-time performance is further compromised when using the sequential strategy for enhancement. To strike a balance between localization accuracy and computational time, we propose a sequence-based place recognition method for LiDAR-Visual fusion, named SeqMamba-MPR. First, we perform joint encoding of LiDAR point clouds and panoramic images using the maximum detection distance, referred to as SMRD.Subsequently, global descriptors are obtained through a single-scan module and a sequence-scan module designed based on a state-space model. The cross-scan design in the single-scan module improves computational efficiency, while the sequence all-combination representation strategy in the sequence-scan module addresses the place recognition problem with trajectory variations in the scene. We conduct experiments on the public dataset NCLT and a self-recorded NEU-Robot dataset, demonstrating that the proposed method achieves a balance between localization recall metrics and real-time performance, outperforming the current state-of-the-art methods.