CrossFusionNet: Robust and Efficient Cross-Modal Place Recognition with Adaptive Feature Fusion

This paper addresses the challenging problem of Image-to-PointCloud (I2P) place recognition, a critical task for autonomous systems hampered by a significant cross-modal gap between RGB images and LiDAR point clouds. We introduce CrossFusionNet, a novel end-to-end framework designed for robust and efficient I2P retrieval through an adaptive feature fusion mechanism. Our approach first aligns Field-of-View and generates dense depth maps from sparse LiDAR for enriched geometric context. A dual-stream encoder extracts multi-scale local features from both RGB images and dense depth maps. The core innovation, the Adaptive Cross-Modal Attention Fusion (ACMF) Module, dynamically integrates these local features with highlevel semantic concepts derived from a lightweight unsupervised semantic clustering layer. This adaptive fusion generates highly discriminative global descriptors, aggregated by independent NetVLAD layers. Trained on the KITTI dataset, CrossFusionNet consistently achieves leading or highly competitive performance in place recognition tasks, demonstrating superior recall across various sequences. Moreover, our model operates with remarkable efficiency, making it suitable for real-time applications. Extensive ablation studies confirm the effectiveness of the ACMF module and semantic clustering, validating our design choices.