Research On Wide-range Map Construction And High-precision Localization Algorithm Based On Multi-sensor Fusion

In recent years,with the rapid development of artificial intelligence technology,autonomous driving has gradually entered people’s vision and become a research hotspot in the field of artificial intelligence.The autonomous driving system mainly includes perception,localization,decision-making,planning,control and other modules.Simultaneous Localization and Mapping(SLAM)is a key technology in the localization module,which is used to build high-precision maps and locate the location of unmanned vehicles.Accurate and robust localization provides basic information for the planning and control module,and also improves the safety of autonomous driving vehicles to a certain extent.The current mainstream SLAM algorithm reduces the accumulated error of Li DAR odometry by loop closure detection.However,when building a map in a large outdoor environment,some scenes do not have loop closure,resulting in continuous accumulated error.Moreover,when localization on the pre-built map,because the map is composed of discrete key frames,when the initial position of vehicle is between the key frames of the map or is not on the mapping trajectory,it is impossible to accurately estimate the vehicle pose directly through the relationship between the current frame and the map keyframes,resulting in global localization failure.In this regard,this paper uses Li DAR,IMU and GNSS as sensors,and proposes a mapping and localization algorithm suitable for outdoor large scenes.The innovative contents of this paper are as follows::(1)Aiming at the problem that the Li DAR odometry keeps accumulating errors when there is no loop closure in the scene,this paper selectively fuses the prior information of GPS to reduce the accumulated errors when the unmanned vehicle is mapping,and can build high-precision point cloud maps even without loop closure.(2)In order to solve the problem that the initial position of the unmanned vehicle deviates from the mapping trajectory,a global descriptor with translation and rotation invariance is proposed in this paper.The descriptor is based on the idea of removing dynamic objects,coordinate system and decentralization,and can be used for the global positioning of subsequent unmanned vehicles.(3)To solve the problem that the global localization is trend to fail when the prior position is approximate,this paper proposes a global localization algorithm based on the global descriptor.The algorithm first performs rough localization through descriptor matching,and then dynamically selects a search range around the rough localization result according to the similarity between the current frame descriptor and the map key frame descriptor,and performs fine localization within this range.(4)After global localization,this paper introduces a tightly coupled factor graph model to fuse the Li DAR,IMU and map prior information to achieve real-time and stable position tracking,and realize lane-level localization on HD map.Finally,experiments are performed in KITTI dataset and real campus data scene,and extensive experimental results verify that the proposed algorithm reaches the level of mainstream algorithms.