Spatial Grids Assisted Multitask Fusing Framework for Incremental Point Cloud Mapping of Large Scenes

Recovering the point cloud maps of large-scale scenes from multi-sensor data is the fundamental building block of numerous 3D related applications. Available SLAM (Simultaneous Localization and Mapping) frameworks require rigorous constraints in data capturing and suffer from heavy workload when facing large-scale scenes. To push the research to further extent, this work proposes an incremental multi-task mapping framework for large-scale scenes under the guidance of spatial grids. (1) This framework first encodes the target scene into spatial grids according to the GNSS (The Global Navigation Satellite System) coordinates, then the mapping task is accomplished via multiple sub-scene SLAM tasks that can be executed independently, and finally the mapped point clouds are incrementally aligned with the assistance of encoded spatial grids to form the whole picture of the large-scale scene. (2) In each sub-scene SLAM task, we develop a new fast loop detection method based on spatial grids and perceptual hashing, which replaces the memory-expensive BoW (bag-of-words) model for image similarity determination, saves the unnecessary computational cost and lifts the localization accuracy. (3) In the aligning part, a new grid-wise merging method is designed for sub-scene point clouds, which improves the robustness and efficiency of point cloud registration and the applicability for large-scale scenes, supporting incremental multi-task mapping. We extensively evaluate the localization accuracy, loop detection efficiency on indoor EuRoC dataset, and multi-machine point cloud mapping on outdoor KITTI dataset. Experimental results reveal that the proposed framework provides a feasible multi-task fusing solution for incremental mapping of large scenes.