It is difficult for gaze detection systems to detect pupils in camera images containing eyeglasses reflections caused by the infrared light sources necessary for gaze detection. In a previous study with a system using two cameras, we proposed a geometrical pupil detection algorithm. However, depending on the positional relationship between the pupil and the reflection, the algorithm was not always effective. In the present study, we arranged three cameras—one at each vertex of a triangle—and composed three camera pairs with different base line directions. We propose several new algorisms, and in the basic algorithm, pupil candidates indicating the suprathreshold angle between two planes, each of which includes the cameras composing a camera pair and the candidate detected by one of the two cameras, are removed from the candidate list. A significant improvement of pupil detection is experimentally demonstrated.
