自然な背景を含むシーンからの身体画像の抽出

抄録

Using a technique of motion measurement basis on video images, limitation of subjects and constraint of its motion are reduced because of no surface markers. In this method, joint angles are estimated from geometric model-matching into a body image, however, it is necessary to extract the body image from a scene including natural background.
In the present paper, we propose a body recognition method using fuzzy reasoning basis on 4 kinds of image features: difference of intensities between an object image and its background image, difference of textures between them, momentary changes of intensities produced by body movements and approximate position and orientation of the geometric model for the model-matching method. These features were calculated on each divided rectangular segment of the image. Shortening the calculating time, we used a hierarchical process as follows. In the first step, the rectangular segments were divided roughly as the first hierarchy. The body segments were recognized globally in this low resolution hierarchy(H16×V16). Then, the segments were divided finer in the next hierarchy, and 8-neighbor segments of the body contours that were recognized in the previous hierarchy were re-recognized locally in this higher resolution hierarchy (H32×V32). Re-recognition of the neighbors on the body contours was continued until the hierarchy resolution was equal to the digital image resolution (H512×V256). Using the fuzzy reasoning method and the hierarchical process, the body image including little holes and little isolated segments could be extracted. It takes 7min per frame by a personal computer.