Eye tracking with statistical learning and sequential Monte Carlo sampling

Many ways have been proposed for eye tracking. These methods are either based on detecting low-level image characteristic or pattern recognition techniques. The first approach is fast yet lack of accuracy. The second approach is accurate but slow. This paper presents a novel method proposed for fast and accurate eye-tracking by a combination of the above methods. A Gaussian mixture model for skin color is built to locate the eyes approximately in a fast speed. Then, probabilistic principal component analysis (PPCA) is applied to confirm the eye location accurately. Sequential Monte Carlo sampling, an enhanced sampling technique, is integrated with the system to further enhance the speed. Experimental results show that it can perform eye tracking accurately with fast speed and robust against the different degree of deformation, orientation gaze and shape of eyes.