关键词: 深度学习；人脸检测；YOLOv8s；实验评估

Abstract: In response to the complex and varied driving scenarios, the large number of parameters in the driver face detection model, and the low recognition accuracy, this paper selects YOLOv8s, which has a small number of parameters, fast detection speed, and high positioning accuracy, as the detector, and designs a driver face detection calculation method based on optimized YOLOv8s to reduce computational complexity and improve detection accuracy. Introducing an attention module in the feature extraction network to increase facial attention, and using a lightweight generalized sequential backward selection (GSBS) module to reduce parameter count. In the feature fusion network, variety of view-group shuffle cross stage partial (VoV-GSCSP) network module and GSBS module are used to reduce the number of parameters, and the upsampling operator content-aware reassembly of features and bidirectional feature pyramid network (BiFPN) structure are used to optimize the feature pyramid and improve the utilization of feature map information. The experimental results show that compared to YOLOv8s, our proposed method not only improves the detection speed of a single image by 6.6 ms, but also increases the average accuracy by 0.5%， meeting the real-time and accuracy requirements of face detection in vehicle systems.

Key words: deep learning; driver face detection; YOLOv8s; experimental evaluation