Webcam-Based Eye Movement Analysis Using CNN

Abstract

Due to its low price, webcam has become one of the most promising sensors with the rapid development of computer vision. However, the accuracies of eye tracking and eye movement analysis are largely limited by the quality of the webcam videos. To solve this issue, a novel eye movement analysis model is proposed based on five eye feature points rather than a single point (such as the iris center). First, a single convolutional neural network (CNN) is trained for eye feature point detection, and five eye feature points are detected for obtaining more useful eye movement information. Subsequently, six types of original time-varying eye movement signals can be constructed by feature points of each frame, which can reduce the dependency of the iris center in low quality videos. Finally, behaviors-CNN can be trained by the time-varying eye movement signals for recognizing different eye movement patterns, which is capable of avoiding the influence of errors from the basic eye movement type detection and artificial eye movement feature construction. To validate the performance, a webcam-based visual activity data set was constructed, which contained almost 0.5 million frames collected from 38 subjects. The experimental results on this database have demonstrated that the proposed model can obtain promising results for natural and convenient eye movement-based applications.