Eye corners tracking for head movement estimation

Recently, video-oculographic gaze tracking has begun to be used in the diagnosis of a wide variety of neurological diseases, such as Parkinson and Alzheimer. For this application, the so-called feature-based methods are used, more precisely, 2D regression-based methods. They use geometrically derived eye features from high-resolution eye images captured by zooming into the user's eyes. The main weakness of these methods is that the head of the user must remain motionless to avoid estimation errors. In some patients, some involuntary movements cannot be avoided and it is necessary to measure them. In this paper, we tackle the measurement of head position as a way to improve the gaze tracking on these precision demanding medical applications. As a first stage, we propose to obtain the eye corners coordinates as a reference point, since they are the most stable points in front of the eyeball and eyelids movements. The problem was handled as a regression problem using a coarse-to-fine cascaded convolutional neural network in order to accurately regress the coordinates of the eye corner. Particularly, with the aim of achieving high precision we cascade two levels of convolutional networks. Finally, we added temporal information to increase accuracy and decrease computation time. The accuracy of the estimation was calculated from the mean square error between the predictions and the ground truth. Subjective performance was also evaluated through video inspection. In both cases, satisfactory results were obtained.