An Experimental Study on Driver Drowsiness Detection System using DL

Driver drowsiness is one of the largest causes of automobile accidents worldwide, resulting in thousands of fatalities and injuries annually. Conventional methods of monitoring drowsiness using vehicle- mounted observation and biological sensors are constrained by their intrusiveness and environmental sensitivity. The developments in the deep learning technology, specifically Convolutional Neural Networks (CNNs), have allowed for the development of strong, non-invasive eye state moni- tors that can detect drowsiness in real-time. This present paper takes into account some of the most popular CNN models, like VGG-16, ResNet- 50, MobileNetV2, and EfficientNet-B0, to determine the top-performing model for real-time drowsiness detection. EfficientNet-B0 is also known as the best choice with its class-leading accuracy-computation ratio. The proposed system employs live video stream, image processing using OpenCV, and Softmax classification for the identification of excessive eye closure and sending early warnings as part of the initiative towards reducing fatigue-related accidents and road safety. The paper also com- pares the performance of different CNN models with regard to different metrics and describes their implications for field deployment. Finally, low-light detection problems, handling real-time behavior, and facial occlusion are addressed with suggestions towards improvement.