3D Indoor Map Construction And Application Based On Visual SLAM

Simultaneous Localization and Mapping(SLAM)is a key technology to realize autonomous navigation of moving objects.With the rapid development of machine vision,Visual SLAM(V-SLAM)has gradually become a hot technology for indoor positioning and navigation.According to the different methods of extracting image feature points,V-SLAM technology can be divided into feature point method and direct method.Feature point method is widely used because it does not require much computation and can save more image information.In this paper,in order to improve the accuracy of V-SLAM indoor positioning and 3D mapping,a more in-depth study of V-SLAM based on feature point method is carried out:First,this paper uses the improved Oriented FAST and Rotated BRIEF(ORB)feature method to process image feature points,and combines the Progressive Sample Consensus(PROSAC)algorithm to eliminate mismatched feature point pairs.The improved orb algorithm adopts quad-tree management to the image feature points extracted by the traditional orb algorithm,and selects the feature points with the largest response value to avoid the phenomenon of clustering of image feature points.Combined with PROSAC algorithm,by comparing the distance ratio between the nearest neighbor and the next nearest neighbor matching points of feature points,select the points with good quality each time to fit the model,so as to eliminate false matching.Experimental results show that the improved orb algorithm can effectively ensure the uniform distribution of image feature points,and PROSAC algorithm can provide accurate feature point matching.Then,based on the traditional loop detection,this paper filters the feature points through the non-maximum suppression method,establishes the image global descriptor,and matches the image global descriptor with the candidate frame global descriptor to further judge the correctness of the loop.At the same time,the Bundle Adjustment(BA)algorithm is used to optimize the camera pose,so as to minimize the accumulated error of the system and improve the accuracy of navigation and 3D mapping.In the RGB-D data set published by the Technical University of Munich(TUM),the experimental results show that the root mean square difference of the absolute trajectory error of the camera is 0.01 ～ 0.02 M,and the root mean square difference of the relative pose error is 0.003 ～ 0.004 m.Finally,this paper builds the simulation platform of RGB-D V-SLAM experiment.The platform is powered by Kinect 2 0 camera,notebook computer and mobile car,which are controlled in Robot Operating System(ROS).In the experiments of data set simulation and loop back positioning,the experimental results show that the system not only has certain advantages in simulation experiments,but also can be well reconstructed and loop back positioning in real-time mapping.There are 40 figures,6 tables and 73 references.