Research On SLAM Of Multi-source Sensor Fusion

In recent years,the development of all walks of life around the world has gradually tended to be automated,and the demand for intelligent robots has increased.Robots are required to make correct decisions in medical care,transportation,autonomous driving and other aspects.As a bridge between robots and external perception,Simultaneous Localization and Mapping technology is particularly important.However,when SLAM is applied to actual scenes,the existence of dynamic objects seriously interferes with the process of laser point cloud registration,reducing the robot’s positioning accuracy.The robot cannot make a correct judgment on the loopback in SLAM when environment cannot be correctly recognized,and the accumulated error cannot be optimized.To solve the above problems,a multi-source sensor fusion SLAM technology is proposed.For the problem of laser odometry positioning deviation caused by dynamic objects,a dynamic point cloud recognition algorithm based on cluster constraints is proposed,and,combined with the process,the traditional feature matching strategy is improved by a combination of coarse registration and fine registration,which constitutes a cluster-based laser odometry design in a dynamic environment.At the same time,the YOLO neural network is used to detect the target of the image,and the recognized target is projected into the laser point cloud through the external parameter matrix.The loopback detection node is optimized,and finally the loopback detection results and the front-end output pose optimized by scan-to-map are optimized to obtain a more accurate real-time pose and generate a real-time point cloud map.The main research contents of the paper are as follows:(1)The dynamic point cloud recognition technology in dynamic environment is studied.First of all,in view of the irregular shape and size of each object in the lidar point cloud,a fast clustering algorithm based on Euler distance is proposed.The Euclidean distance-based clustering of ordinal point clouds can quickly segment the point clouds corresponding to independent and non-connected semantic individuals.Then,a dynamic point cloud recognition algorithm based on cluster constraints is proposed for the problem that dynamic objects will affect the positioning accuracy.The algorithm calculates the cluster center of each point cloud cluster,transforms it with the matrix pre-estimated by the IMU,compares it with the point cloud cluster class center at the previous moment,quickly establishes the point cloud cluster class matching relationship and removes the dynamic point cloud in the point cloud by setting a threshold.(2)The point cloud registration technology in complex environment is studied.Aiming at the long iteration time of a single ICP registration process,a point cloud registration scheme based on cluster constraints is proposes.First,the large grid where the point cloud cluster is located is replaced by the self-segmented small grid in the NDT algorithm.The grid performs rough registration on two point clouds whose relative relationship is completely unknown to obtain the initial pose transformation matrix,and then iterates each feature point through the ICP algorithm to improve the accuracy of point cloud registration.The combination improves the accuracy and speed of traditional feature matching strategies.(3)The loop closure detection technology of multimodal fusion is studied.Aiming at the problem that the traditional loop closure detection scheme is constrained by distance,a global descriptor construction method based on the height information of the environment is proposed.The neural network performs target detection on the corresponding camera image,uses the external parameter matrix to project the detected target into the point cloud,completes the semantic annotation of the point cloud,and then encodes and calculates the global descriptor through the completed semantic annotation to generate a Dimension operator,and with the help of KD tree’s fast processing characteristics of high-dimensional data,the task of reducing potential candidate key frames is completed,thereby optimizing the operation efficiency of loop closure detection nodes.Finally,the minimum similarity value is obtained by fine-tuning the column offset to determine whether there is a loopback.This method greatly reduces the time complexity of loop closure detection,and verifies the existence of loop closures through semantic information in the environment,avoiding the problem of the traditional odometerbased loop closure detection method being affected by accumulated errors,and can effectively reduce the time and cost of loop closure detection.Mismatch rate.(4)Through the KITTI dataset and the designed indoor environment localization experiment,dynamic environment experiment and outdoor localization experiment,the effectiveness of the cluster constraint-based laser odometry and the multi-modal fusion SLAM back-end optimization in a dynamic environment are respectively verified.The experimental results show that the algorithm designed in the paper has high accuracy and can obtain high-precision positioning results in a complex experimental environment.