Real-time 3D human pose recovery from a single depth image using principal direction analysis

3-D human pose recovery using nonrigid point set registration and body part tracking of depth data

“…Figure 1 describes the steps of our processes. First, our system used a human T-pose depth silhouette, yielding a body part-labeled map by pixel-supervised classification via random forests [15] for initialization. After initialization, each human depth silhouette that was represented as the set of 3-D points was matched to the previous silhouette via point set registration to obtain point correspondences.…”

“…We asked a subject to make a T-pose, obtained a depth silhouette, and created the body part-labeled map. To label body parts on the human depth silhouette of the T-pose, we used a pixel-wise supervised classification via trained random forests [15]: the training needed only a small synthetic T-pose database for labeling body parts. The human depth silhouette and its labeled map with fifteen labeled parts of a T-pose are shown in Fig.…”

“…For comparisons against the conventional methods, we have evaluated the performance of our proposed methodology by comparing against the conventional methods [9,15]. For comparison against mean shift-based method [9], we reproduced the human pose recognition and used our synthetic DB in [15] to train RFs.…”

“…For comparison against mean shift-based method [9], we reproduced the human pose recognition and used our synthetic DB in [15] to train RFs. We evaluated the mean shift method through quantitative assessments on the synthetic dataset.…”

“…The quantitative assessment result with the same synthetic DB of our proposed method and the method [9] is presented in Table 1. The average [15] which is based on body-part labeling as introduced in the supervised learning-based method. The mean reconstruction error of the PDA method at eight key joints was 7.10° compared to 6.26° of our method.…”