Recognizing Human Postures and Poses in Monocular Still Images

In this paper, person detection with simultaneous or subsequent human body posture recognition is achieved using parts-based models, since the search space for typical poses is much smaller than the kinematics space. Posture recovery is carried out by detecting the human body, its posture and orientation at the same time. Since features of different human postures can be expected to have some shared subspace against the non-person class, detection and classification simultaneously is tenable. Contrary to many related efforts, we focus on postures that cannot be easily distinguished after segmentation by their aspect ratio or silhouette, but rather require a texture-based feature vector. The approaches presented do not rely on explicit models nor on labeling individual body parts. Both the detection and classification are performed in one pass on the image, where the score of the detection is an ensemble of votes from parts patches.