SDF-TAR: Parallel Tracking and Refinement in RGB-D Data using Volumetric Registration

This paper introduces SDF-TAR: a real-time SLAM system based on volumetric registration in RGB-D data. While the camera is tracked online on the GPU, the most recently estimated poses are jointly refined on the CPU. We perform registration by aligning the data in limited-extent volumes anchored at salient 3D locations. This strategy permits efficient tracking on the GPU. Furthermore, the small memory load of the partial volumes allows for pose refinement to be done concurrently on the CPU. This refinement is performed over batches of a fixed number of frames, which are jointly optimized until the next batch becomes available. Thus drift is reduced during online operation, eliminating the need for any posterior processing. Evaluating on two public benchmarks, we demonstrate improved rotational motion estimation and higher reconstruction precision than related methods.