Publication,
Research,
Docs and Info,
KAMEDA
Motion estimation which aims to handle deformable objects is one of the important research in image understanding. Among the deformable objects, a human body has a significant value to be estimated because it is very useful to recognize a user's intension and instruction.
Recognition of motions of a human body is a challenging problem in computer vision. Rohr[4], Bharatkumar[3], Attwood[5] have contributed to solve this problem. Many researches use a stick or cylinder model and edges in images. However, human bodies are not stick in a strict sense. Moreover, edges in images usually contains not only useful ones but also many meaningless ones.
In this paper, we propose a human motion estimation method using a difference image sequence. A difference image is obtained from continuous two images which are taken by a fixed camera. Our method is based on a model matching method. A human model consists of 15 nodes each of which represents a part of a human body and has a joint attribute. In this paper, motion estimation is defined to estimate the joint angles in the model. In each frame of an image sequence, our model matching algorithm takes a node of the model one by one. We classify a status of the nodes into three modes; moving mode, stationary mode, and occlusion mode. A node is labeled with one of three modes by considering the relationship between the difference image and the projected region of the node on the image plane. We calculate joint angles according to the modes of the nodes at each frame. If a node is in the occlusion mode, the images cannot be used. To cope with such situation, we introduce inertia constraint such that the parts in a human body are generally rotated at a constant speed.
We assume that an image sequence does not include any moving object except for a human body and the camera calibration is done in advance and that the initial pose of the human body model is obtained from an other method such as [1][2].