GPS Assisted Divide-and-conquer SLAM Framework For Large-scale Scenes

In the past thirty years,computer vision has experienced rapid development.With the advancement of technology,SLAM(Simultaneous Localization and Mapping)has affected people’s lives,which is a key supporting technology for devices applied to all kinds of applications,such as the robots in small-scale scenes,unmanned vehicles in medium-scale and large-scale scenes,autonomous drone navigation,and epidemic AR/VR devices.With the gradual expansion of SLAM application scenes,the traditional SLAM algorithms suffer from the problems of large calculation volume,huge memory consumption and large cumulative drift when facing large-scale scene mapping.They cannot provide a reliable solution.As the current hardware equipment is equipped with multiple sensors,it can be considered to integrate multiple sensor data,especially GPS information with global reference significance,so that the SLAM system can still ensure accurate and rapid positioning and mapping in large-scale scenes.This paper proposes a novel GPS information-assisted large-scale scene SLAM framework in divide and conquer principle.Taking GPS information as a reference,the large-scale scene SLAM problem is divided into sub-scenario SLAM tasks that can be running independently by multiple machines,and finally the reconstructed sub-scene point clouds are automatically aligned to complete the rapid localizing and mapping of large-scale scenes.The framework consists of three main modules:(1)Space division and coding: A dynamic space division and coding strategy for large-scale scenes is proposed.Guided by GPS information,the scene space is recursively divided into grids using an octree structure,and the gird codes are dynamically generated.By giving global coding information to the image keyframes that fall into the spatial grid,global guidance and constraints are applied for subsequent subscene SLAM and fast point cloud alignment;(2)Sub-scene SLAM: Fuse image and IMU data to realizes the visual odometer to estimate the robot poses,and then a fast loop detection algorithm based on space division/coding is proposed to improve the accuracy and efficiency of relocation and back-end pose optimization.The space codes are firstly employed for image screening before loop detection,then the perceptual hashing method is applied to compute and compare the image similarity.Hence,the loading and maintenance of the bagof-words model is avoided,the loop detection process and calculation are simplified,and finally the trajectory estimation accuracy of SLAM tasks is improved.Benefiting from the global constraints of space grids,the sub-scene SLAM tasks can be performed independently by multiple machines.(3)Large-scale scene point cloud fusion: A fast and automatic aligning method for multi-region reconstruction point clouds is proposed,which integrates GPS global information into the restored scene 3D point cloud,and constrains the rapid aligning of sparse point clouds of sub-scenes.Owing to GPS-assisted point cloud alignment,the large-scale scene reconstruction can be accomplished in the way of regional growth.Subsequently,the point cloud data optimization registration method adapted to the spatial gridding and coding is designed,and the grids of the overlapping region are used as guidance to complete the bottom-up point cloud fine registration of neighboring areas efficiently and accurately,which improves the accuracy and applicability of the framework for large-scale scene reconstruction.Finally,the designed divide-and-conquer SLAM framework and three important algorithms proposed in this thesis are evaluated on public datasets KITTI and Eu Ro C.The analysis are carried out from the aspects of loop detection efficiency,trajectory estimation accuracy,point cloud fusion efficiency and accuracy,etc.,to show its positioning and mapping performance for large-scale scenarios.The thesis also compares the performance with other baseline algorithms to verify the efficiency and feasibility of the framework.The results show that the GPS-assisted large-scale scene divide-and-conquer SLAM framework proposed in this thesis can quickly obtain accurate trajectory estimation and scene sparse point clouds,support multi-person and multi-machine parallel,seamless data connection,and incremental completion of scene SLAM and 3D point cloud reconstruction,providing solutions for largescale scene positioning and mapping.