Blinking Eyes Detection using Convolutional Neural Network on Video Data
F. M. Sigit, Eko Mulyanto Yuniarno, Reza Fuad Rachmadi, Ahmad Afan Zaini · 2020
The drowsiness conditions at human body can affect to changing some parts of body's behaviour, such as changing behaviour of eyes, mouths, and brains. When people in drowsy conditions, distance between upper eyelid and lower eyelid at their eyes will be shorter than normal behaviour, mouths will be a little more opened rather than usual, and brains will produce very low frequencies of brain's signals. This changing behaviour in some parts of bodies gives a benefit to detect drowsy conditions when some people do not give concern to this. Different behaviour of faces and eyes during drowsy conditions and normal conditions; decreasing distance between upper eyelid and lower eyelid is a worth idea to detect drowsiness, this is a basic idea in this research, we want to build a machine learning to detect blinking eyes based on images running in realtime. In this research, we collect images of faces and eyes to build a dataset and separate this dataset to two label categories based on our target classifications, these labels are ”opened eyes” and ”closed eyes”. There are 3 different datasets in this research, first dataset contains 6000 images, second dataset contains 8000 images, and third dataset contains 10000 images of faces and eyes, each of those datasets is collected from one sample person, one sample person is me (author). There is a little thing different at 10000 image dataset compared by those two existing datasets particularly at closed eyes class category. In closed eyes class category at 10000 image dataset contains two variations of closed eyes images, there are 4000 perfectly closed eyes images and 1000 half-closed eyes images. All of those datasets are trained to convolutional neural network (CNN) so we have 3 different pretrained CNNs. Those three pretrained CNNs are tested to detect blinking eyes of samples running in realtime. There are 11 differences of samples, one sample is me (author) and 10 other samples are from other people. From this test, we get the conclusions that the highest success in detect blinking if those pre-trained CNNs are tested to detect blinking eyes from sampel face of me (author) is exactly placed at center in front of frame/camera. Rate of success in this detection is 0.95 every 20 detection.