3D Reconstruction Of Indoor SLAM Based On Point-line Feature And Hybrid Closed-loop Detection

Simultaneous Localization and Mapping(SLAM)is the key technology in the field of robotics,which has a wide range of application.In the structured environment,there are plenty of line features and point features in the image,which can complement each other to structure the basic architecture of the images.In this thesis,the RGB-D SLAM three-dimensional reconstruction scheme based on point-line features and hybrid closed-loop detection is constructed by depth camera.The purpose is to improve the accuracy of positioning and reconstruction,and to enhance the robustness of the whole scheme.The main contents of this thesis are as follows:Firstly,the improved pose estimation algorithm based on point-line feature is proposed.The point-line feature is correlated with data,and the error model of point-line feature is established.Combining the sparse direct method with the feature point method,this thesis presents a coarse-to-fine pose estimation method which bases on the point-line feature.ORB feature points are extracted from RGB images,and the current camera pose is coarse estimated by sparse direct method.Then,line features are extracted.Based on the former pose,the graph optimization model is constructed by point-line features,and the camera pose is precisely estimated.Then,the closed-loop detection method is improved.On the basis of Randomized Ferns algorithm,the bag of word is introduced to make up for the defect of Randomized Ferns'sensitivity to light.The selection strategy of candidate loop closing frames is improved,which code the two images using Randomized Ferns algorithm,and calculate the Hamming distance of them.Then,the Hamming distance is normalized,and the similarity is calculated by combining the word bag vectors of the two images.By doing this,the accuracy of closed-loop detection can be improved while the image information is fully utilized,and the globally consistent camera pose and maps can be obtained,provide more accurate maps for re-location and navigation.Finally,this thesis establishes the 3D reconstruction scheme of indoor SLAM based on point-line features and hybrid closed-loop detection.The scheme includes tracking thread,local mapping construction,closed-loop detection thread and the reconstruction of three-dimensional scene based on surfel.ICP algorithm and color information of image are used to optimize the global pose again when reconstructing with surfel.A series of comparative experiments on TUM data sets show that the proposed algorithm improves the accuracy of camera trajectory to a certain extent and shows the effect of Closed-loop Detection and three-dimensional reconstruction.