Keyframe Based RGB-D Simultaneous Localization And Dense Mapping

Simultaneous localization and dense mapping is an important subject in computer vi-sion.Dense 3D information of the scene find it's applications in many fields.For example,a dense 3D map can assist the robots to get rid of obstacles;In AR,handling of occlusion will be easy with the help of dense scene 3D structure.In recent years,with the popularity of consumer-grade RGB-D cameras such as Microsoft Kinect,Asus Xtion,and the rapid perfor-mance increase of GPU hardware,more and more attention has been paid to simultaneous localization and dense mapping of scenes using RGB-D cameras.In RGB-D based simultaneous localization and dense mapping,the estimation of camera pose and the reconstruction of globally consistent 3D models are the two most important issues.In this regard,this paper does an in-depth study and builds a keyframe-based RGB-D simultaneous localization and dense mapping system,which can reconstruct a globally consistent 3D model in real time.Based on this,an AR system is developed,which can effectively handle occlusions and generate more realistic experience.The main contributions of this paper are as follows:1)RGB-D alignment is introduced to the feature-based RKD-SLAM[1]system proposed by Liu,which improves it's robustness effectively in weakly textures scenes.2)In large-scale RGB-D reconstruction,to correct the model efficiently when the camera pose is updated due to loop closure,keyframe based re-integration method is proposed,which can adjust the model quickly and effectively when the camera pose is updated.3)Based on the above-mentioned 3D reconstruction method,an augmented reality sys-tem is implemented,which can effectively handle the occlusion between real and virtual scenes.