Graph Neural Network for Human Body Orientation Estimation by 2D Skeleton

Service Robot has been increasingly gaining attention from hospitality to tourism businesses especially in the Human Robot Interaction section. Human pose estimation has added to increase the understanding from human to robot. The famous estimation tools are MediaPipe Pose, MMPose, and Detectron2. However, these tools did not come with orientation prediction. Orientation is important in predicting the intention of humans, especially in “Pedestrian Research” and “Handover Task Research”. Therefore, in this paper, we present a simple method add-on to the existing human pose estimation tools to predict the human body facing direction. First, we extract the key points from the existing tools and then use them to compute a set of nodes and edges features to form a graph. Afterwards, we use Edge-Conditioned Convolution (ECC) for graph level prediction to predict the orientation of the human pose. Experiments conducted on the TUD dataset and show that the proposed GNN outperforms previous works on 8 discrete orientation classification.