Underdetermined approach to real-time face tracking and recognition

In a real-time face tracking and recognition system proposed by Oka and Shakunaga, an optimum weighted average of registered images are estimated and the weights are used for face identification and shape inference.Although their method works well even when a target face changes pose in photometric changes, both the person identification and expression recognition could not be robustly solved at the same time because of a capacity problem. This paper proposes underdetermined approach to solve the capacity problem. Although a single underdetermined system often results in some performance reduction, parallel implementation can remarkably improve the performance. Experimental results showed that the proposed method successfully worked for 10-person discrimination when 10 expressions were registered for each person even when an image sequence included many face motion, expression, photometric change and occasional occlusions.