Research On Algorithms For Point Line Visual SLAM And Millimeter Wave Radar SLAM

Simultaneous Localization and Mapping(SLAM)is one of the key technologies of the robot application system.For vision sensors in low-light and low-texture scenes,and millimeter-wave radar sensors with less effective data and more noise,the pointline feature is the basic feature for sensing the environment and positioning and mapping.Based on point and line features in vision and millimeter-wave radar point clouds,combined with an inertial measurement unit.This paper conducts research on point-line vision-inertial navigation SLAM and millimeter-wave radar SLAM algorithms to achieve accurate positioning and mapping in multiple scenarios.The main research contents include:1.Optimize the quality of visual SLAM feature extraction.Aiming at the problem that the accuracy of point feature extraction in the environment is not high enough,an iterative method is proposed to find sub-pixel corner points.And constrain the iterative corner point coordinates based on distance and boundary.In the line feature extraction,some short lines are judged and eliminated based on the length of the line segment,and the line features are supplemented by multi-angle homogenization.2.The efficiency and accuracy of visual SLAM nonlinear optimization are improved.For the situation that the same line segment is repeatedly recognized during line feature extraction,it is judged based on the angle,distance and length between line segments,and the repeated line features are merged.During the nonlinear optimization of the system,the visual pose information and other information are optimized step by step according to the difference in the amount of variable data to be optimized.3.The feature optimization of millimeter-wave radar SLAM point cloud.In view of the small amount of effective data in the millimeter-wave radar SLAM system,the iterative closest point method is used to match the radar point cloud,and the matching relationship between the front and rear frame point clouds is constrained.For the situation where there are many noise points in the system,the random sampling consensus algorithm is used to filter the outliers.For frame jumps that may occur during system scanning,a jump optimization function is established to screen effective key frames to ensure global positioning accuracy.The global pose is constrained using loop closure detection for radar positioning trajectory shifts over time.In the visual SLAM experiment,the accuracy is verified using the EuRoC public dataset and the dataset recorded by the handheld IMU camera.Compared with algorithms VINS-Mono,PL-VIO and PL-VINS with better performance,the positioning accuracy and system processing efficiency have been significantly improved.In the millimeter-wave radar SLAM experiment,the MulRan public dataset was used to verify the improved accuracy of the algorithm.Use the Boreas dataset to compare the positioning performance of vision and millimeter-wave radar SLAM under the same road segment at the same time.Using the nuScenes data set to conduct experiments on the fusion effect of vision and millimeter-wave radar sensors at the feature level,and analyze the feasibility of the fusion of the two in the SLAM system.