Spherical scan context: a global descriptor in the form of a third-order tensor for loop closure detection

Loop closure detection is a crucial component of laser SLAM (Simultaneous Localization and Mapping) systems, used to eliminate long-term accumulated errors. However, describing the discrete point cloud data collected by laser SLAM using simple geometric structures is more challenging than using the rich image information in visual SLAM. This makes location identification based on point clouds particularly difficult. Most existing work does not consider the correlation between the spatial and intensity data of the point cloud or relies on prior training, resulting in models that do not generalize well. In contrast, we propose a method to characterize historical locations by combining spatial and intensity data from point clouds. We exploit rotational invariance in sphere geometry features and perform fast filtering and matching of previous historical data with a two-stage search algorithm. Our experiments on the KITTI dataset demonstrate that our method has significantly improved performance.