Cross transformer for LiDAR-based loop closure detection

Loop closure detection, also known as place recognition, a key component of simultaneous localization and mapping (SLAM) systems, aims to recognize previously visited locations and reduce the accumulated drift error caused by odometry. Current vision-based methods are susceptible to variations in illumination and perspective, limiting their generalization ability and robustness. Thus, in this paper, we propose CrossT-Net (Cross Transformer Net), a novel cross-attention based loop closure detection network for LiDAR. CrossT-Net directly estimates the similarity between two frames by leveraging multi-class information maps, including range, intensity, and normal maps, to comprehensively characterize environmental features. A Siamese Encoder Net with shared parameters extracts frame features, and a Cross Transformer module captures intra-frame context and inter-frame correlations through self-attention and cross-attention mechanisms. In the final stage, an Overlap Estimation Module predicts the point cloud overlap between two frames. Experimental results on several benchmark datasets demonstrate that our proposed method outperforms existing methods in precision and recall, and exhibits strong generalization performance in different road environments. The implementation of our approach is available at: https://github.com/Bryan-ZhengRui/CrossT-Net_Pytorch.