Adaptive Eye Aspect Ratio Technique for Drowsiness Detection System

Drowsiness Detection System (DDS) is a system that can detect driver fatigue based on the state of the subject's facial features, body pose, behaviour, and inner features in order to give a warning alarm to prevent the driver from an accident. Several approaches can be used to build DDS, one of which is using face landmark information to estimate the Eye Aspect Ratio (EAR). However, this technique is only good at detecting frontal faces. To handle this problem, we developed a modified EAR-based DDS system with additional information on the direction vector of the face. To identify the driver condition, we use an Attention Mesh model to extract the face landmark of the driver. Then by using the EAR technique, we can estimate the eye state of the driver. Then we use the projection constant and machine learning technique to process both the EAR information and the direction vector of the face to estimate better eye conditions. The experimental results show that our model is the most balanced in accuracy and speed. The highest result comes from the gradient boosting classifier and extra trees classifier tested on the CEW dataset and the dataset we collected on the driver's environment in daytime and nighttime conditions. Meanwhile, in terms of computation speed, the traditional EAR thresholding method is still faster than our method, up to around 10x faster. Even so, our method is still fast enough to achieve real-time detection.