Design And Implementation Of Tightly Coupled SLAM Algorithm Based On Multi-sensor

With the development of computer and automotive technology,autonomous driving is closer to our life.Autonomous driving technology mainly contains several modules such as perception of environment,mapping,localization and vehicle control.The Simultaneous Localization and Mapping algorithm(SLAM)can estimate and optimize the pose of the vehicle,and build a high-precision point cloud map by collecting real-time data from multiple sensors.This ultimately provides assurance for the localization and mapping of autonomous driving.However,the information acquired by a single sensor is limited,and relying on single sensor data for pose estimation will lead large errors.In contrast,the approach of multi-sensor fusion can fully utilize the advantages of different sensors.Compared to the approach of optimizing individual sensors separately in a loosely-coupled system,tight-coupling can jointly compensate for information from multiple sensors,thereby improving the overall stability of the SLAM framework.Therefore,most researchers consider multi-sensor tightly-coupled SLAM to improve mapping performance.With the increasing driving trajectory of the vehicle,the errors generated by pose propagation in existing multi-sensor tightly-coupled SLAM algorithms will accumulate,eventually leading to map drift.However,loop closure detection can establish loop constraints between current frame and historical frame to decrease accumulated errors.Therefore,research on loop closure detection is crucial.To solve the problem that the loop closure detection module in the tightly-coupled framework is difficult to correctly identify loops,this paper proposes a robust loop closure detection scheme and improves the multi-sensor tightly-coupled SLAM algorithm.This paper’s work is as follows:(1)To address the issue of the problem of loop closure detection failure caused by rotation and translation of the vehicle in the same scene,this paper redefines the origin and main direction of the point cloud,unifying the coordinate system and thereby reducing the impact of sensor perspective changes on scene description.(2)This paper designs a 3D global descriptor with twice projection to improve the mechanism of loop closure detection.The descriptor utilizes a multidimensional encoding method and integrates the height and intensity information of the point cloud,thereby enhancing its recognition and discrimination abilities for the scene.(3)This paper utilizes the property of three-dimensional descriptors that can be projected from multiple directions,and applies side-view and top-view projections.Based on the projected results,a weighted similarity calculation method is designed to reduce the false positive rate in loop closure detection.(4)This paper integrates the above loop closure detection method into a multisensor tightly-coupled SLAM framework,introducing the generalized iterative closest point method in the back-end for point cloud registration,thereby constructing loop closure pose constraints.Additionally,the algorithm constructs a factor graph in a tightly-coupled manner using information from Li DAR,IMU,and the above loop closure constraints,which is used to optimize the vehicle’s pose.In addition,different platforms are used to collect data in both campus and offroad environments.Subsequently,loop closure detection and SLAM algorithm experiments are conducted on both publicly available datasets and self-collected datasets to verify that the proposed algorithm performs well in real-world scenarios.