Visual Place Recognition (VPR) constitutes a pivotal task in the domains of computer vision and robotics. Prevailing VPR methods predominantly employ RGB-based features for query image retrieval and correspondence establishment. Nevertheless, such unimodal visual representations exhibit inherent susceptibility to environmental variations, inevitably degrading method precision. To address this problem, we propose a robust VPR framework integrating RGB and depth modalities. The architecture employs a coarse-to-fine paradigm, where global retrieval of top-N candidate images is performed using fused multimodal features, followed by a geometric verification of these candidates leveraging depth information. A Discrete Wavelet Transform Fusion (DWTF) module is proposed to generate robust multimodal global descriptors by effectively combining RGB and depth data using discrete wavelet transform. Furthermore, we introduce a Spiking Neuron Graph Matching (SNGM) module, which extracts geometric structure and spatial distance from depth data and employs graph matching for accurate depth feature correspondence. Extensive experiments on several VPR benchmarks demonstrate that our method achieves state-of-the-art performance while maintaining the best accuracy–efficiency trade-off.

Accurate and real-time localization plays a pivotal role in low-cost vehicle autonomous driving applications. Cross-modal visual localization in LiDAR-built maps is a feasible solution. For limiting drift, it is of utmost importance to address the challenge of balancing priormap compression while preserving essential geometric structureinformation and achieving real-time data association. This paper proposes a novel method, dubbed PVMLoc, for monocular visual localization in prior voxel map. Voxels store semantics, while surfels are extracted from semantic point cloud. This method employs hashed voxel indices stored in an unordered map, enabling efficient dynamic loading of submaps and data association based on raycast operations. Then, according to the semantic guidance, local ground manifold constraints and surfel constraints are constructed to optimize the visual local map and the pose of the alignment prior map, respectively. Finally, PVMLoc evaluates the performance of the cross-modal localization system in the prior voxel map on the KITTI and self-constructed NEU localization dataset, thereby demonstrating the advantages of each module. Comparative analysis against state-of-the-art algorithms demonstrates that the method achieves competitive localization accuracy.