Research On 3D Reconstruction Algorithm Based On Binocular Vision

In this paper,Kinect depth camera and Simultaneous Localization and Mapping(SLAM)are used to reconstruct indoor scenes.Aiming at the demand of this technology,this paper mainly establishes the 3D-reconstruction system for indoor scenes from the aspects of image feature extraction and matching algorithm,camera pose estimation,back-end optimization and closed-loop detection of visual SLAM.This paper mainly completed the following tasks:(1)In the aspect of feature matching,based on Affine Oriented FAST and Rotated BRIEF(AORB)feature matching algorithm,proposes a matching algorithm of multiple strategy combinations.Firstly,feature extraction is performed with AORB,rotation-aware BRIEF(rBRIEF)establishes descriptors,reduces the number of initial matching points,and improves matching efficiency.Then,the constructed descriptors are matched by the method of cross matching and K-Nearest Neighbor(KNN).Finally,the matching minimum distance is combined with the Progressive Sample Consensus(PROSAC)algorithm to eliminate the mismatch points,improve the matching precision and reduce the matching time.(2)In the optimization of pose estimation,the PROSAC algorithm is used to calculate Perspective-n-Point(PnP)and the motion transformation of adjacnt frames.The PROSAC only uses a few random points to calculate PnP and determine the inner points,but it is vulnerable to noise interference.In order to improve the method,PROSAC solution is used as the initial value,PnP is used to find the spatial position transformation relationship between the connected frames,and then non-linear optimization is used to optimize the pose transformation matrix to get the optimal solution,so as to improve the accuracy of camera motion pose estimation.(3)In the aspect of back-end optimization and closed-loop detection,weights are introduced in key frame selection,frame similarity is calculated with Bag of words(BoW)visual dictionary,inter-frame constraints are added,loop detection is performed to eliminate accumulated error,and pose maps are optimized by general graph optimization(g2o)algorithm.So as to realize the optimization and updating of 3D-reconstruction point cloud maps.Finally,the trajectory of the camera and the point cloud maps of the reconstructed scenes are obtained.The experiments in this paper are based on Ubuntu16.04+KDevelop,which realize the functions of each module.The experimental results show that the proposed algorithm can quickly construct a globally consistent indoor 3D-reconstruction environment map.The experimental results basically meet the expected standard,and the results are accurate and reliable.