3D Human Pose Machine with a ToF Sensor using Pre-trained Convolutional Neural Networks

This paper proposes a new system for estimating 3D human pose with a ToF sensor. In the proposed system, a single low-cost ToF sensor captures depth data of human pose. Then, a new clean imaging converter for the ToF sensor transforms the depth data corrupted with sensor errors and zero values into a clean 3D image data. Finally, a deep learning predictor based on the convolutional neural networks pre-trained with millions of 2D image data of human pose, not 3D ones, estimates the 3D human pose from the clean 3D image data. Experimental results show that the proposed system shows the good performance comparable to the commercial system in terms of 3D human pose estimation accuracy.