Coarse to fine global RGB-D frames registration for precise indoor 3D model reconstruction

Abstract

The conventional approach to register two or more RGB-D frames produced from low cost depth sensors, such as KINECT and Structure Sensor, applies SIFT matched points between color images along with corresponding depth from the depth images. This is known as RGB-D SLAM. This method depends on ICP concept to refine the sensor pose after estimating it from SIFT depth points. In this research, we propose a new registration method and a new description function to add line feature matching in RGB-D frame registration. The qualitative and quantitative assessments of the proposed procedure show a significant improvement in 3D model quality and precision with the proposed new registration method.