Research On Indoor Dense Mapping Method Based On Improved ORB-SLAM2 Algorithm Based On RGB-

Simultaneous Localization and Mapping(SLAM)is a key technology for realizing truly fully autonomous mobile robots,which can be divided into laser SLAM and vision SLAM.Vision SLAM technology is favored by universities and enterprises because of its low cost and rich information,etc.The emergence of RGBD camera makes the vision SLAM technology based on this sensor become a hot spot for robot intelligence research.The main objective of this paper is to improve the ORB-SLAM2 system based on RGB-D cameras,to study the indoor dense map building algorithm,and finally to realize a more accurate and widely used visual SLAM system.The main research of this paper includes the following three parts.(1)Proposed improved AKAZE(Accelerated KAZE)algorithmThis paper uses the improved AKAZE algorithm to replace the ORB(Oriented FAST and Rotated BRIEF)feature extraction algorithm in the ORB-SLAM2 system to improve the quality of features.The improved AKAZE uses AKAZE algorithm to extract key points of grayscale images,BRIEF(Binary Robust Independent Elementary Features)algorithm to obtain descriptors,and introduces Intensity Centroid and convolution kernel function to enhance the robustness of descriptors when using BRIEF for computation.To verify the effectiveness of this algorithm,fifteen sets of data from the TUM dataset are compared in this paper,in which the average repeatability rate of the improved AKAZE algorithm is improved by 0.3974 compared with the ORB algorithm.(2)Proposed improved Gaussian mixture model algorithmThe process of the improved Gaussian mixture model algorithm is to first match the features extracted by the improved AKAZE algorithm using the violent matching algorithm in two frames of the ORB-SLAM2 system,then recover the color information of the grayscale map using the pseudo-color algorithm,after which the Gaussian mixture model is used to segment all pixels in the R channel of the image,and finally retain the features in the same segmented region.In this paper,the proposed improved AKAZE algorithm is fused with the improved Gaussian mixture model algorithm and validated using fifteen sets of data from the TUM dataset,and its average matching accuracy is improved by 0.1867 compared with the ORB algorithm,which proves that the matching method can be used more effectively for subsequent vision tasks.(3)Proposed a new thread for constructing octomapThe new thread based on ORB-SLAM2 constructs a dense point cloud map using the key frames obtained from the visual odometry,and then converts the map into an octomap after voxel filtering.Octomap has the advantages of low memory usage,easy maintenance and storage,and can also be used for positioning,navigation,obstacle avoidance,and 3D reconstruction.The improved ORB-SLAM2 system is validated using four different datasets,and the root mean square error is reduced by 0.0037 on average in the calculation of relative positional error considering both translation and rotation,the number of point clouds in the point cloud map is reduced by 2226429.75 on average after voxel filtering.