A lightweight method for fatigue driving detection based on facial analysis

Fatigue driving detection is of significant importance for ensuring driving safety. Aiming at the issues of low accuracy and slow inference speed in existing methods for real-world driving scenarios, a novel lightweight method for fatigue driving detection based on facial analysis is proposed. Firstly, for each frame in the input driving monitoring video, the driver’s eye and mouth regions of interest (ROI) are extracted using face detection and facial key points detection algorithms which avoids the need for manual parameter tuning in traditional region extraction methods and can adapt to various head postures. Secondly, an efficient state recognition network for facial ROI called SRNet-FR is designed to recognize the open or closed state of the extracted eye and mouth region images, in which the Ghost module is used to make the model lightweight and a 3D non-parametric attention mechanism called SimAM is employed to improve the classification accuracy without introducing additional parameters. Finally, fatigue-related parameters are calculated based on the driver’s eye and mouth states of each frame, and the driver’s fatigue state is determined according to the established fatigue state determination rule. Experimental results show that the trained SRNet-FR achieves a high accuracy of 99.03% for eye and mouth state recognition under various head postures and in various driving scenarios which has a parameter amount of only 0.61M and a computation amount of just 16.45MFLOPs. The proposed method achieves an accuracy of 97.3% for fatigue driving detection on test videos which outperforms other compared methods and has an inference time of only 41 milliseconds for a single frame on the CPU platform.