Real-time Monocular Dense Mapping of Small Scenes with ORB Features

In this paper, we present a new solution for real-time dense mapping with single moving camera. Our monocular dense mapping system remove needs of extra power supply and limitation in indoor scenarios almost all active sensors require. Our system present robust camera pose estimation and can recover from tracking failure. By leveraging ORB-SLAM [1], high accuracy 6DoF camera pose and sparse feature point map are first provided. Then we propose a motion stereo method for every frame's depth map generation in real-time. The depth map is then integrated into volumetric scene model. For each frame, the implicit represented surface is extracted and information in other parts of system is presented in GUI for live visual feedback. Our system provides more robust tracking performance and provide compelling result visually comparable to RGBD systems in ORB-SLAM in small scenes. Our system is both efficient and easy to implement. We detail algorithms used in system pipeline and specific configurations of our system in this paper. Our CUDA based implementation runs near $25\mathrm {H}\mathrm {z}$ on mainstream laptop.