DESIGN OF A BLINK DETECTION SYSTEM USING COMPUTER VISION AND TRANSFER LEARNING

• Main Author: Thomas Widjaja, James
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/79388
• Institution: Institut Teknologi Bandung

Long shift work duration and inadequate supervision contribute to an increased risk of incidents and accidents due to fatigue. Thus, a real-time fatigue measurement is needed to monitor and measure the levels of workers’ fatigue during work. One of the techniques that have been known to be reliable is a fatigue measurement technique that focus on non-visual characteristics. However, this technique is considered as an intrusive technique. Therefore, the other fatigue measurement technique that focus on visual characteristics is more preferred. Ocular indicators are considered as one of the most important fatigue indicators and thus can be applied for measuring fatigue in realtime due to its non-intrusive characteristics. Furthermore, blink parameters such as blink frequency and blink duration are considered as the most important and informative parameters for measuring fatigue. This study aims to design an algorithm and system prototype for blink detection then test the validity of the system-generated blink parameters. This system is designed using computer vision and machine-learning classifier based on convolutional neural network (CNN), MobileNet, with transfer learning method in order to classify the eye state. The use of computer vision in blink detection is presumed to be influenced by external conditions such as fluctuating lighting intensity and vibrations. Two video recordings consist of two different external conditions are used to test the presumption of different external conditions influencing the system performance in blink detection. System performance is measured using confusion matrix and the validity of the blink parameters are tested using statistical tests. In conclusion, the proposed algorithm and system prototype are capable of detecting blink and generate blink parameters such as blink frequency, blink duration, microsleep and percentage of eyelid closure (PERCLOS), however the system is still considered insufficient for real-time fatigue measurement based on performance reasons. Furthermore, the system-generated blink parameters are proven reliable for measuring fatigue, but only under controlled external condition.