REAL-TIME AND DEEP LEARNING-BASED FATIGUE DETECTION FOR DRIVERS

Among the causes of traffic accidents, driver errors are in the first place. Driver faults are generally considered to be situations such as drunk driving and excessive speeding. However, sleep-deprived and tired driving are also among the leading causes of driver faults. Driving while feeling sleepy and fatigued; Effects such as slow reaction time, decreased awareness, and inability to focus occur. Considering this situation, it is understood that driving while sleepy and tired is at least as dangerous as driving under the influence of alcohol. In this study, a system that works in real-time inside the vehicle constantly monitors the driver and works with high accuracy is proposed. This system is deep learning based and low cost. In the study, the driver's eye and mouth movements were analyzed to determine normal, yawning and fatigue. A data set has been created for this. The data set consists of videos taken at different times and in different ways from 129 volunteers. Videos shot in different formats, quality and sizes were collected, and turned into a single format. Grayscale, tilt addition, blurring, variability addition, noise addition, image brightness change, color vividness change, perspective change, sizing, and position change were added to the photographs that make up the data set. With these additions, the error that may occur due to any distortion that may occur from the camera is minimized. Thus, the accuracy rate in the detection process with images taken from the camera in real-time has been increased. At the same time, a new data set specific to the study was prepared. YOLOv5, YOLOv6, YOLOv7, and YOLOv8 architectures were used in the study. The newest and most used architectural results in the literature are compared. As a result of the study, a 98.15% accuracy rate was obtained in YOLOv8 architecture. It is aimed that the study will be highly effective in preventing traffic accidents.