3D Human Pose Estimation from RGB-D Images Using Deep Learning Method

In human activity recognition system, detecting the human and estimating the pose of 2D or 3D human correctly is critical issue. In this paper, we propose a novel approach for 3D human pose estimation from RGB-D CCTV images based on a deep learning approach. By using the RGB-D model rather than the conventional RGB image which has a limitation in detecting object due to the lack of topological information, we can resolve the self-occlusion problem and improve the object detection ratio from efficiently. Subsequently, the position of a human joint is localized with a Convolutional Neural Network (CNN) from the detected person. In this phase, we utilize CPM (Convolutional Pose Machine), to generate belief maps to predict the positions of key-point representing human body parts and estimate 2D human pose by detected key-points. In final stage, we estimate 3D human pose from the 2D joint information based on Deep Neural Network (DNN). From the experiment, we prove that the proposed method detects human objects robustly in occlusion and the estimated 3D human pose are very accurate comparing the previously introduced methods. As for the future work, the estimated 3D human pose will be used for human activity recognition.