Part Detector Based Human Pose Estimation in Images and Videos

Human pose estimation is an essential issue in computer vision area since it has many applications such as human activity analysis, human computer interaction and visual surveillance. In this paper, 2D human estimation issue in monocular images and videos is addressed. The observation model and the inference method are improved based on part based graph inference method. A rotation invariant edge field feature is designed and based on which a Boosting classifier is learnt as the observation model. The human pose estimation is done with a particle based belief propagation inference method. Experiments show the effectiveness and the speed of the proposed method.