YOMO-ECA: A Lightweight Multi-Facial Feature Fatigue Driving Detection Method

Driver fatigue is one of the key causes of road traffic accidents. It is crucial to ensure driver safety and improve road safety. To address the current problems of driver fatigue detection, such as complex detection, low accuracy, and susceptibility to the influence of complex environments, a YOLO-based algorithm for detecting the state of driver's eyes and mouth, YOMO-ECA, is proposed. Firstly, the Bottleneck part of the lightweight MobileNetV3 model is used to reconstruct the backbone of YOLOv8, achieving model lightweighting. Then, the Efficient Channel Attention(ECA) mechanism is used to replace the original Squeeze and Excitation (SE) mechanism of MobilentV3, the ECA is also introduced after each of the four C2f in the neck part of YOLOv8, to enhance the feature representation capability of the model and capture the complex correlations and dependencies in the contextual information. Moreover, the Focal-EIOU Loss is used to replace the original CIOU to reduce the impact of label class imbalance. On this basis, Percentage of Eye Closure Time (PERCLOS) and Percentage of Open Mouth (POM) are used as the fatigue evaluation criteria, and the two are fused as the fatigue evaluation indexes, which overcomes the limitation of single feature fatigue detection. The experimental results show that compared with the original YOLOv8n, the YOMO-ECA algorithm improves the mAP by 1.1%, reduces the Params by 26.8%, and reduces the FLOPs by 61.7%. The method proposed in this paper is able to maintain high detection accuracy and accurately recognize driver fatigue state while being lightweight.