Odometry Optimization And Loop Detection Based On Multi-beam LiDAR Mapping

With the concepts of artificial intelligence known by more people and the advent of autonomous cars and related projects,the autonomous driving has become one of the hottest topics at the moment.Due to its high precision,high reliability and low external interference,LiDAR has become an indispensable sensor in autonomous systems.The dense point cloud map based on LiDAR is also commonly used in positioning methods.However,the traditional algorithms cost a long time with slow speed,and they cannot detect a loop closure quickly in a large environment during mapping process.This thesis is based on the relatively cheap 16-beam LiDAR and IMU and other auxiliary sensors.The research content is to carry out faster loop detection while quickly obtaining the unmanned vehicle's positions during the construction of the mapping process,and improve the efficiency of the construction of mapping in a SLAM system.The mapping system based on 16-beam LiDAR can be divided into three parts: the acquisition of the front-end pose estimation,the feature extraction for loop detection,the loop search and match.The preprocessing of the point cloud data can eliminate the point cloud distortion.By establishing a suitable motion model,the motion compensation for each point can be derived according to its angle.After the point cloud data conforming to the physical world,by matching selected appropriate feature points of each frame,the corresponding poses can be calculated.Aiming at the translational cumulative error and the obvious rotational drift problem existing in the point cloud matching process,our method established a motion model for the correction process of point cloud distortion,the motion of the car is predicted by acquiring the IMU and speed sensor with higher frequency.Transformation between poses can be obtained by merging transformation from point cloud matching result through Extended Kalman Filter,then the system can get a more precise estimation of the poses.For the feature extraction part of point cloud in the mapping process,this thesis proposes a simple and fast method for extracting the poles like features as the basis of loop detection.The system combines the sparse point cloud of successive frames to obtain a denser point cloud for cluster segmentation,and then hierarchically projects each cluster to the grid to obtain the number of occupied grids.Then,the data continuity in the Z direction of the cluster is judged whether it belongs to the poles feature,and finally the obtained poles features are divided into corresponding features and spatial positions for storage.For loop detection,this thesis adopts a loop detection strategy based on frontend pose estimation.Under the premise of obtaining a more accurate pose estimation,the system matches the feature distribution of the current moment with the searched historical feature distribution.Through the matching between feature distributions,the rotation translation relationship between the current pose and the historical pose can be obtained,and the corresponding constraints are obtained.Then,the transformation between each pose and the transformation relationship of the global constraint are input into the backend general graph optimizer,its output,the optimized result,is the final poses estimation in the whole mapping process.