Human3D Body Pose Estimation Based On Multi-scale Histograms Of Oriented Gradients

The human pose estimation is the hot research topic in the computer vision. By the humanpose estimation, the computer can automatically get the human pose information from the humanbody image or video sequence. Then he human pose information can help the computeranalyzing and understanding the human behavior, which greatly improved the intelligent degreeof the machine vision system.Because of the human body posture changes, the human skeletal structure differences,perspective changes, different dressing, illumination changes and many other factors, whichleads to difficult in3D human pose estimation. Especially in the monocular environment it existsself shielding, and loss of depth information and other uncertain factors, the human poseestimation is more difficult.In this paper, we propose a new method to estimate3D human body pose from monocularimages. In the proposed method, image features are represented by the multi-scale histograms oforiented gradients (Multi-scale HOG) and3D body poses are represented by3D limb angles.Relevance Vector Machine is used to infer body poses from the image features.Compared with the commonly feature descriptors, such as the context of shape, and scaleinvariant feature transform method, the histogram of oriented gradients contain more humanappearance information. We use the multi-scale histogram of oriented gradients to eliminate theeffect of the big gradient amplitude in the local histogram of oriented gradients. The multi-scalehistogram of oriented gradients is a method to calculate the information of image features inmulti-scale space. It can express the complete information of the body contours and the internaltexture details in different scales. Furthermore, to reduce the disturbance of the background andenhance the robustness of our method, background is removed by silhouette feature in theMulti-scale HOG computational procedure of our method.In the most of the human pose estimation algorithms,3D coordinates of human joints areused to represent a human body pose. The disadvantages of3D joints based pose representationlie in that the dimension is large and sensitive to the perspective change in monocularenvironment. Instead, in the proposed algorithm, we use3D limb angles to represent a humanbody pose which can overcome these disadvantages.The experimental results show that our method can accurately estimate the human pose in amonocular environment. Meanwhile, in the case of the camera angle changes, our method canalso accurately estimated the3D limb angles. This means that our method is robust for thecamera angles changes.