Research On Lidar SLAM Technology Based On Visual Information Assistance

With the advent of 5G "Internet of Everything" era,there is an increasing demand for intelligent and smart mobile service platforms.Simultaneous Localization and Mapping(SLAM)technology,as the core technology and prerequisite basis for intelligent activities such as tracking,navigation and path planning,has been a hot research topic for universities and enterprises at home and abroad in recent years,and has been widely applied in the fields of autonomous driving,logistics and distribution,and community services.SLAM solutions that are more mature include lidar SLAM and visual SLAM.Because of the high accuracy and the stable performance of LIDAR sensors,lidar SLAM is currently the mainstream solution used in industry.As various smart mobile platform industries continue to change,the situations in which SLAM is used are becoming more and more complex.Relying only on the scanning information of LIDAR sensors may not be comprehensive enough to meet the requirements of accuracy and robustness in complex environments,so SLAM solutions utilizing other sensors for assistance or fusion are significant research directions for the future development of this technology field.Among them,visual cameras have become one of the most widely used sensors due to their low price and rich radiation information.The image acquired by the visual camera is 2D,while the laser point cloud of LIDAR is 3D,so the geometric information of LIDAR is richer,but the image radiation information acquired by the camera is richer,thus the two can be combined organically to avoid the shortcomings.Therefore,we make full use of the advantages of both,study and implement the lidar SLAM algorithm using visual information to assist in loop detection,and build a more robust and reliable SLAM system.The main works are as follows:(1)Studying the fundamental theory of monocular visual camera and LIDAR fusion.Firstly,the basic theory of monocular visual camera and LIDAR fusion was described,including coordinate transformation and sensor model;then,the nonlinear optimization problem and its common optimization methods were introduced,and the nonlinear least squares problem was described;Finally,based on the sensor characteristics of LIDAR and monocular camera,focusing on accurate and fast state estimation and sensor information fusion scheme,a complete SLAM system including efficient and accurate front-end odometer,and reliable and robust back-end loop detection map optimization was constructed.(2)A LIDAR odometer scheme with high-frequency direct matching method and low-frequency feature matching method was proposed to address the contradiction between accuracy and real-time performance of traditional SLAM front-end odometer.Firstly,the classification basis of the laser point cloud scan matching algorithm was introduced;secondly,the GICP matching localization algorithm based on the idea of voxelization was introduced,the nearest neighbor search method of voxelized GICP was improved,and the key frame matching method across multiple frames(Scan2Key)was adopted,a LIDAR localization scheme based on improved GICP matching was designed,and the lowfrequency odometry strategy based on the feature point method was proposed,which can further improve the state estimation accuracy and effectively balance the relationship between the state estimation accuracy and systematic real-time;finally,the effectiveness of the improved voxelized GICP algorithm in this paper was verified through offline point cloud alignment experiments,and the state estimation accuracy and realtime performance of the front-end odometer scheme in this paper were tested and evaluated by the KITTI dataset and the data collected from the SLAM platform.(3)A reliable and robust back-end technology solution for loop detection and global graph optimization was proposed by making full use of the visual sensor characteristics.Firstly,we explained the importance of loop detection for SLAM system,combined the respective characteristics of LIDAR and visual cameras,designed a fusion method using visual word pocket detection,laser point cloud verification,and introduced a binary dictionary to optimize the efficiency of word pocket search;secondly,we introduced the construction and optimization of global pose graph,based on the improved front-end LIDAR odometer,built a complete back-end optimization scheme based on monocular vision-assisted improved lidar SLAM.finally,we tested the accuracy of back-end loop detection,as well as the cumulative error and map consistency after global map optimization using KITTI dataset and the data collected from the SLAM platform,and developed a narrative and insightful analysis of the test results.