ADVANCES IN EYE BLINK DETECTION: EVALUATING CONVOLUTIONAL NEURAL NETWORK BASED BLINK DETECTION ON REAL WORLD DATA

Abstract

Subjects with varying races, wearing glasses, hair in the way of their eyes and periodically moving faces makes it difficult for convolutional neural networks (CNN) to detect eye blinks in scenarios such as detecting driver fatigue. However, if the accuracy of the CNN is too low, then using it for eye blink detection in life like scenarios would not be advisable because the CNN would be incorrectly classifying to many times. To increase the accuracy of the CNN researchers have tried varying race of subjects and situational profiles. The research goal in this paper is to show a similar method to increase the accuracy of a premade convolutional neural network that detects eye blink by varying the training data with different races and situational profiles. The races tested included Caucasian, Hispanic and Asian. Situational profiles included eyeglasses, multiple people in frame, and hair being in the way of the eye. The training data also incorporated frames of video data; this meant the subject periodically moved their face in different directions. Being able to detect eye blink is an important task by itself; therefore, with an accurate CNN we could further the field in eye blink detection. Additionally, I have shown that training with a variety of races, wearing glasses, hair in the way of their eyes and periodically moving faces increased the accuracy (Figures 5,8). However, even though I found an increase of accuracy, many mysteries still remain.