Human Pose Recognition under Cloth-like Objects from Depth Images using a Synthetic Image Dataset with Cloth Simulation

This paper proposes a method of human pose recognition when the body is largely covered by cloth-like objects such as blankets. Such a recognition is useful for robotic monitoring of the elderly and the disabled. Human pose recognition under cloth-like object is challenging due to a large variety of the shape of covering objects. Since we would like to use depth images for addressing privacy and illumination issues, it further makes the problem difficult. In this paper, we utilize computer graphics tools including cloth simulation for generating a synthetic dataset, which is then used for training a deep neural network for body parts segmentation. We achieved around 90% accuracy in synthetic data and show the effectiveness of simulating cloth-like objects in data generation. We also applied it to real data and examined the results for identifying remaining issues.