5. Gesture Analysis Based on Head Position

Based on our face tracker, we have also developed an algorithm for analyzing of the absolute position of the faces. We assume that the camera position is fixed, so that the 2D motion of the face and its features in the video sequence reflects the actual motion of the head.

This algorithm detects shaking and nodding of the head, different viewing scenarios (such as near/distant, uprigh, lying down) and periods of high & low activity. This analysis is useful to sense not only explicit events such as yes- and no-like gestures, but many others clues as to the user's state of mind. These could potentially improve machine interfaces by adapting their behaviors in response to the user's conduct.

In order to detect head shaking and nodding, a time-window of 2 seconds is analyzed every second to detect sinusoidal motion independently on each axis (horizontal and vertical). Then, a rule-based system is used to decide whether the person's head is shaking, nodding, staying stationary or moving erratically. The head position is taken to be the center of the outer eye corners and the motion over the analyzed time-window is normalized using the average rotation angle and distance between the eyes.