LiDAR-SLAM loop closure detection based on multi-scale point cloud feature transformer

Loop closure detection is an important part of simultaneous localization and mapping (SLAM), used to detect and correct map and trajectory drift issues caused by accumulated errors. For the LiDAR-SLAM system, the sparsity and disorder of point clouds make position recognition based on 3D point clouds more challenging. In recent years, many deep learning-based closed-loop detection algorithms have been proposed. However, due to the sparsity of point clouds, current deep learning algorithms often convert point clouds into histograms or depth maps and then process them using deep learning algorithms, undoubtedly causing information loss. In this paper, we propose a closed-loop detection method based on multi-scale point cloud features transformer, which introduces multi-scale point cloud feature extraction and transformer global context modeling. We use voxel sparse convolution to obtain features of original point clouds at different resolutions and establish contextual relationships between features at different resolutions using the transformer network to achieve multi-scale feature fusion, and then obtain global descriptors. The obtained global descriptors can be used not only for closed-loop detection but also for front-end registration to address the challenges of point cloud processing in the SLAM system, especially in enhancing global modeling capabilities and reducing information loss. Our method directly processes point cloud data and integrates multi-scale point cloud feature information, which can better adapt to the characteristics of LiDAR-SLAM systems, improving the accuracy and robustness of localization and map construction, thus having broad application prospects in the field of measurement. We evaluated our method on multiple sequences of the KITTI and KITTI-360 datasets, each containing more than 5000 frames of point clouds, and extensively evaluated on a self-collected dataset of over 3.6 km. The experimental results show that our method achieves an accuracy of over 80% on multiple datasets and demonstrates superior performance in different environments.