Dense RGB-D SLAM with Planes Detection and Mapping

Liang Pan, Pengfei Wang, Jiawei Cao, Chee-Meng Chew

Abstract: Observing the absence of predominant planar features in most previous RGB-D simultaneous localization and mapping (SLAM) systems, we introduce a dense RGB-D SLAM, which meanwhile detects and visualizes large planes in the reconstructed indoor scenes. The major challenges are threefold. Firstly, large indoor planes are usually partially observed. Moreover, loop-closure problems should not undermine those detected and reconstructed 3D planes. At last, the efficiency, especially the processing time analysis, is always a major concern in SLAM systems. To unravel these problems, we detect plane segments in each new observed depth map. The detected new and old planes are matched and updated in a frame-to-model fashion. Hence, our plane detection results are directly related to the reconstructed 3D scenes, which eliminates the influence of loop closures. We enhance the efficiency of our system by considering those detected integral plane segments instead of individual points during camera motion tracking. Furthermore, our system accelerates all the processes by heavily applying parallel computations. Experimental results demonstrate that our system can densely reconstruct 3D scenes with detected planes, which also achieves near real-time property.