Low-Cost Depth Camera Pose Tracking for Mobile Platforms

The KinectFusion algorithm is now used routinely to reconstruct dense 3D surfaces at real-time frame rates using a commodity depth camera. To achieve robust pose estimation, the method conducts the frame-to-model tracking during camera tracking that must inevitably accompany the memory-bound, GPU-assisted volumetric computations for the model manipulation, to which mobile processors are often more vulnerable than PC-based processors. In this paper, we present an effective camera-tracking method that is based on the computationally lighter frame-to-frame tracking method. This method's tendency toward rapid accumulation of pose estimation errors is suppressed effectively via a predictor-corrector technique. By removing the costly volumetric computations from the pose estimation process, our camera tracking system becomes more efficient in terms of both time and space complexity, offering a compact implementation of depth sensor-based camera tracking on low-end platforms such as mobile devices in addition to high-end PCs.