Answers are grounded in this paper and cite their sources.
Place recognition plays a vital role in applications such as autonomous driving, drone-based geo-localization, and serves as a key component of Simultaneous Localization and Mapping (SLAM) systems. However, challenges such as viewpoint changes, illumination variations, and weak textures hinder the performance of place recognition, especially in unstructured environments. To address these issues, this paper presents a multimodal approach that fuses visual and point cloud data. Two dedicated branches are used for encoding visual and point cloud data, enabling efficient local feature extraction from each modality and providing dual-modal features for fusion to obtain comprehensive global representations. To enhance multimodal features alignment, the extrinsics between the camera and LiDAR are exploited. Building on this, a semanticaware alignment-based pruning strategy is introduced to suppress redundant information and focus on meaningful correspondences. Meanwhile, a transformer-based hybrid attention mechanism, combining self-attention and cross-attention, is employed to extract discriminative global features for robust representation. Following coarse global feature matching, the top-k candidates are re-ranked through a local feature weighting strategy that integrates both visual and point cloud local features. Our approach is validated on two public datasets, demonstrating strong performance across diverse environments. Experimental results further show that our method consistently outperforms existing baselines in these settings, highlighting its potential to enhance performance of SLAM as well as its effectiveness in broader place recognition applications.
No takes yet. Share an insight, caveat, or question.
Shi et al. (2025) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: