Gesture recognition based on a path searching method in weighted graphs for focusing on important action

In the previous studies on automatic gesture recognition for a multimodal human-computer interface, the focus was on relatively large motions such as a wide swing of the arm; it did not apply to subtle differences in local motions such as the hand gestures seen in sign language. Yet, comprehensively speaking, there are extremely similar actions in gestures and local actions play an important role in differentiating those actions. In this paper, we propose a matching method focused on important partial action features that could be obtained by dividing the optical flow of actions. With the matching method, it became possible to recognize actions that have important meanings in their local actions. We describe that the method was realized by applying a graph path-planning method to the action matching and reducing the cost of the graph edges corresponding to important features. Also presented are the results of an experiment on the recognition of similar sign words.