PSV-LCD: Point cloud Surface Variation based Loop Closure Detection

Loop closure detection remains an active research field in simultaneous localization and mapping. In this study, we propose a loop closure detection method based on layered point cloud surface variation. A local map is constructed by accumulating consecutive keyframes using odometry data. Three layers of points are extracted from each local map. Each layer is partitioned into regular 2D grids to form a bird's eye view image where each pixel value is the surface variation of the points that fall within the corresponding grid. The partition of the local map into layers exposes different structural patterns in the environment. We detect ORB features on each layer and, using the bag-of-words concept, we identify loop closure candidates. Layer-to-layer matching and pose computation are performed between the current local map and each loop candidate, using RANSAC-based geometric verification. For each layer-to-layer matching, we require a minimum number of inlier features. We also require that the poses obtained from the layer matches agree. This multi-step validation results in robust loop closure detection. The proposed approach is tested in different environments, including forest environments, and achieves high precision and high recall results. The source code is publicly available at https://github.com/IssoufGit/psv_bev_loop.