Articulated Pose Estimation with Global Body Orientation

Human pose estimation in monocular still image is a hotspot and difficult issue in the computer vision field, and it has important applicable to many novel area. However, the detecting rate is still poor due to many factors as variety of pose and clothing, occlusion, clustering background etc. A pose estimation approach combining deformable part model and full body model is presented to solve the problem of mistaking part orientation caused by the detecting error for body location. We use full body model to augment the spatial position of full body with head-shoulder detector. We find that the whole body location is benefit to capture human body parts. The experimental results on PARSE dataset show that our approach outperforms many past studies for several body parts.