Estimating Human Body Orientation using Skeletons and Extreme Gradient Boosting

Abstract

Human body orientation is a valuable information for social robots due to its use in social path planning and person (or group) approaching. Multi-sensory is an alternative for orientation estimation but it is not available for in-the-wild applications. To estimate orientation using only a single image, several computer vision techniques have demonstrated insufficient accuracy. We propose combining 2D skeleton information with extreme gradient boosting algorithm to detect orientation. We obtain person's skeleton using the OpenPose deep architecture, and extract its distances and angle features. These attributes are used to train a gradient boosting learning system by XGBoost. To evaluate predictions considering real situations based on a single camera, the TUD Multiview Pedestrian dataset is used. We compared the proposed approach against various state-of-the-art methods and our results indicate better classification performance. Furthermore, we prove that our method is viable for body orientation estimation on real-life scenarios by presenting case studies on simulated scenes.