Simultaneous Localization And Local Dense Map Construction Based On Keypoint Method

Simultaneous localization and dense map construction is the key for mobile robots to realize environment exploration and navigation.How to improve the accuracy of camera localization in environment with illumination and viewpoint changes,how to estimate the dense depth map with the existing image information,how to express the dense map and fuse each observation are the key problems to be solved in the simultaneous localization and dense map construction.On the basis of previous research works,the mobile robot simultaneous localization and dense map construction are studied in this thesis.Aiming at the problem of the keypoint matching of camera localization in environments with various changes,and the estimation of dense depth map,as well as the expression of dense map and the fusion of observations,some improved algorithms and schemes are proposed to improve accuracy of camera localization and dense map construction simultaneously.This dissertation has the following main research contents:1.For camera localization based on keypoint method,keypoint matching is easy to fail in environments with illumination and viewpoint changes,leading to the failure of data association.In this dissertation,since convolution feature has a certain robustness in environments with illumination and viewpoint changes,the Conv Net landmark is proposed as an intermediate representation to improve the quality of the putative keypoint matches,which can improve the robustness of the keypoint matching to some extent.2.For dense map construction,dense depth map is the key input information for dense map construction.Inspired by the sparse space points generated by ORB-SLAM2 system and the powerful regression ability of convolutional neural network,a dense depth map estimation framework based on residual network is proposed in this dissertation,which can effectively generate dense depth map from sparse depth of ORB-SLAM2 and the corresponding single image.3.For the dense map building,this dissertation puts forward the local dense map construction framework based on keyframe.Specifically,this framework fuse each observation through Truncated Signed Distance Function(TSDF)fusion and ORB-SLAM2 camera tracking results,and maintains the global consistency of the dense map through online correction.Furthermore,the proposed framework regularizes the built dense map,and maintains the size of the dense map through sliding window,so as to build the dense map quickly and effectively.