Dense Frame-to-Model SLAM with an RGB-D Camera

In this paper, a dense frame-to-model Simultaneous Localization And Mapping (SLAM) with an RGB-D camera is proposed, which achieves a more accurate trajectory in contrast to traditional frame-to-model methods. In the frontend, dense photometric information and geometric information are combined to perform a more robust tracking. In the backend, we add volume to loop closure detection to reject false loop. A novel volume-camera pose graph is proposed to effectively reduce drift. Experimental results on some RGB-D SLAM datasets show a reduction of global trajectory error by 18.60% in comparison to Kinituous, 84.43% in comparison to Kinfu.