Vision-Based Driver Drowsiness and Distraction Detection through Behavioral Indicators of Fatigue

Driver drowsiness is a leading cause of road accidents which emphasizes the need for early detection to improve safety. Fatigue impairs a driver's ability to make safe decisions, endangering both the driver and others on the road. The development of predictive models for drowsiness detection through behavioral indicators has proven to be more cost-effective and resource efficient than physiological or vehicular-based approaches. This study examines the effectiveness of a driver drowsiness detection system by analyzing visual behavioral indicators of fatigue, such as eye closure, yawning, head pose, and eye gaze, to assess distraction levels. Initially, OpenCV and Dlib's Shape Predictor is employed to locate and analyze the facial region of interest and computed measures of drowsiness including Eye Aspect Ratio (EAR), Mouth Aspect Ratio (MAR), head pose estimation using the Perspective-n-point (PnP) algorithm through nose projection, and eye gaze direction through pupil detection. Subsequently, the study explores MediaPipe Face Mesh in combination with the Face Landmarker model as an alternative to Dlib's 68-point landmarks shape predictor for building more robust solution. The MediaPipe solution utilizes blendshape coefficients as features such as eyeBlinkLeft, eyeBlinkRight, JawOpen, Chin, and Nose landmarks for drowsiness detection providing an alternative solution. The study found that MediaPipe Landmarker and Face Mesh outperformed Dlib Shape predictor in drowsy frame detection, offering stronger fitting of facial features and better performance in diverse backgrounds and frames with glass appearances. MediaPipe solution achieved accuracy of 960% in detecting drowsy states, while the Dlib Shape Predictor method achieved an accuracy of 87%.