Affiliation:
1. School of Optics and Photonics Beijing Institute of Technology Beijing China
2. Key Laboratory of Photoelectronic Imaging Technology and System Ministry of Education Beijing China
Abstract
AbstractLiDAR‐based 3D place recognition is an essential component of simultaneous localization and mapping systems in multi‐scene robotic applications. However, extracting discriminative and generalizable global descriptors of point clouds is still an open issue due to the insufficient use of the information contained in the LiDAR scans in existing approaches. In this paper, we propose a novel spatial‐temporal point cloud encoding network for multiple scenes, dubbed STM‐Net, to fully fuse the multi‐view spatial information and temporal information of LiDAR point clouds. Specifically, we first develop a spatial feature encoding module consisting of the single‐view transformer and multi‐view transformer. The module learns the correlation both within a single view and between two views by utilizing the multi‐layer range images generated by spherical projection and multi‐layer bird's eye view images generated by top‐down projection. Then in the temporal feature encoding module, we exploit the temporal transformer to mine the temporal information in the sequential point clouds, and a NetVLAD layer is applied to aggregate features and generate sub‐descriptors. Furthermore, we use a GeM pooling layer to fuse more information along the time dimension for the final global descriptors. Extensive experiments conducted on unmanned ground/surface vehicles with different LiDAR configurations indicate that our method (1) achieves superior place recognition performance than state‐of‐the‐art algorithms, (2) generalizes well to diverse sceneries, (3) is robust to viewpoint changes, (4) can operate in real‐time, demonstrating the effectiveness and satisfactory capability of the proposed approach and highlighting its promising applications in multi‐scene place recognition tasks.