Simultaneous Localization And Mapping Based On RGB-D Camera Key Technology Research

With the development of machine vision technology,vision-based real-time environment modeling technology plays an important role in telemedicine,artificial intelligence,autonomous driving and other fields.It has received extensive attention from academics and industrial researchers and has achieved fruitful results.In recent years,with the continuous development of new depth cameras,real-time 3D environment modeling technology based on active depth acquisition has become a hot spot.However,for lightweight and miniaturized devices,there are still poor accuracy,poor robustness,and maps in real-time 3D reconstruction.Occupy a lot of storage space and other issues.This article mainly carried out the following work in response to the above problems:Firstly,it analyzes the domestic and foreign development status of visual threedimensional real-time reconstruction technology,researches the basic theory of machine vision,the system framework of robot coordinated positioning and mapping technology,and compares and analyzes the advantages and disadvantages of different local feature matching algorithms.The improvement of the core algorithm lays the foundation.Secondly,aiming at the problems of low efficiency,poor accuracy,and poor robustness in the point cloud registration process,a visual odometer that integrates the SURF-FREAK feature point error model is designed.Based on the ICP algorithm,the image matching technology is fused,and the SURF feature extraction and the FREAK feature description are combined to complete the rough matching,which provides the initial value for the fine registration optimization link.In order to improve the robustness of the fine registration process,a depth camera error model is added to the feature point spatial position error model.At the same time,a new distance cost function is used to solve the problem that Euclidean distance cannot measure the feature point error model.The accuracy of the feature points and the reduction of the cumulative error of the visual odometer,after fine matching,the Kalman filter is used to filter the location errors of the feature points to improve the location of the feature points and the depth camera to reduce the error.Experiments show that this algorithm has a positioning error of no more than 30 mm for the depth camera visual odometry problem.The real-time performance is better than the mainstream Pn P-BA algorithm,and the real-time performance and accuracy are higher than the traditional SVD-ICP algorithm.Thirdly,based on the completion of the registration and the splicing of the point cloud maps,a multi-region octree segmentation algorithm is designed to solve the problems of point cloud data redundancy and occupy a large amount of storage space.On the basis of ensuring the accuracy of the environment model,reduce the consumption of computer memory resources during the reconstruction process.Firstly,through pre-regional division and parallel computing by computer,each cube area is divided into octrees at the same time;then,the map data is compressed using the storage structure of octrees to save storage space while ensuring the accuracy of the indoor environment model.Experiments show that based on the multi-region octree segmentation algorithm,compared with the point cloud map,the voxel fusion map eliminates redundant data and compresses 98.5% of the data.Finally,build a ROS operating platform based on the Linux system,use TUM and NYU standard data sets to test the visual mileage calculation method and parallel octree voxel model in the real-time 3D reconstruction of this paper.The experiment shows that its positioning accuracy and robust performance meet Actual demand.