| With the increasing requirements for scene model refinement in indoor 3D modeling,virtual reality,augmented reality,3D games,etc.,real-time dense reconstruction technology based on RGB-D data has received more and more attention.The current RGB-D-based real-time dense reconstruction algorithm mainly uses feature matching or direct registration of image information to obtain camera pose,and obtains a consistent global scene model through closed-loop detection and pose optimization.The specific research work and innovations are as follows:Firstly,several real-time dense reconstruction algorithms based on RGB-D data are introduced and analyzed in detail.These real-time dense reconstruction algorithms generally include trajectory tracking,closed-loop detection,pose optimization,and mapping.This paper introduces the main processes of these algorithms,analyzes the advantages and disadvantages of the algorithms,and provides the basis for the improvement of the algorithms.Secondly,a dense reconstruction algorithm based on ORB-SLAM2 is proposed.On the basis of ORB-SLAM2,the scene dense reconstruction module is added to obtain realtime dense reconstruction results.The image information is fused into the scene model after the camera trajectory is obtained,and the scene model is adjusted according to the position of the camera and the reconstructed sparse point cloud optimization after camera pose optimization.At the same time,in the closed-loop detection part,the consistency constraint is used to obtain high-reliability feature matching results,and the accuracy of the closedloop is effectively verified,and a more reliable closed-loop result is obtained,thereby effectively improving the precision of real-time dense reconstruction.Thirdly,in view of the problems existing in the traditional ElasticFusion algorithm,three improved strategies are proposed to effectively improve the accuracy of dense reconstruction of the scene.A weighted method of joint registration in adaptive adjustment is proposed to solve the problem that the fixed weight in the traditional ElasticFusion algorithm unable adapt to the change of geometric information and texture information in different scenes.The method of re-estimating the camera pose of the error tracking frame is mainly to solve the problem that the traditional ElasticFusion algorithm has some frame tracking errors during the tracking process.This paper proposes to increase the registration between key frames and adjust the camera pose and scene model through the registration result to alleviate the drift phenomenon of the traditional ElasticFusion algorithm over time in non-closed loop situations,and builds effective connection relations between key frames to reduce the degree of trajectory drift.Experimental results show that the algorithm has good performance. |
