Kameda, Research, Documents and Info

Introduction

Recognizing a pose of articulated objects including human beings could be very useful. It could be applied to develop a non-contact input device in exercise physiology, free hand pointing device in man machine communication, and could also be a starting point towards body-language recognition.

Several model-matching methods have been proposed where the strategy depends too much upon the object of estimation, and the information about the model is often only implicitly implemented in their algorithm[1,2,3]. If the model is implicitly defined in the algorithm, it is difficult to know which part of the algorithm uses the model information. In such a case, the method could hardly be applied to other target objects. We perfectly separate the model definition from the algorithm.

In the model-matching method, the selection of image clues is an important issue. Much research has applied edges as the image clue[1,2,3]. However, edge detection has a tendency to extract a large amount of noise and it is hard to find out the desired edges. On the contrary, silhouette extraction has the advantage that it rarely involves noise though a silhouette has less information than edges. We use the silhouette information in our model.

We explain our articulated object model and the contour based method in Section 2, and apply it to a real-life object in Section 3. Experimental results are shown in each section.