Abstract: The paper proposes a distracted driver detection method based on the improved Yolov8. The article designs a weighted bidirectional feature pyramid network (BiFPN) with a small object detection layer and a novel connection structure on the basis of Yolov8. In the backbone network, an efficient multi-scale attention (EMA) module is introduced to enhance the learning and extraction of key information. A new C2f structure is proposed to integrate dynamic convolution and attention mechanisms into C2f. The improved model achieves a 4.6% increase in accuracy while reducing the parameter volume by 27.1%, thus reducing the model size to a certain extent and improving computational efficiency. Finally, generalization experiments on public datasets show that all evaluation metrics are better than those of the original model. Therefore, the research in the article provides a more efficient and accurate solution for distracted driving behavior detection, offering strong support for road safety and traffic regulation.

Key words: Deep learning; Attention mechanism; Driver behavior recognition; Target detection algorithm

摘要： 文章提出了一种基于改进 Yolov8 的驾驶员分心检测方法，在 Yolov8 的基础上设计了 一种包含小目标检测层和新型连接结构的加权双向特征金字塔网络（BiFPN）。在骨干网络中， 引入了高效多尺度注意力（EMA）机制模块，以加强对关键信息的学习和提取。提出了一种 新型的 C2f 结构，将动态卷积、注意力机制融入到 C2f 中。改进模型的精度提高了 4.6%，同 时参数量降低了 27.1%，一定程度上减小了模型的规模，提高了计算效率。最后在公开数据 集上进行泛化实验，各项评价指标皆优于原模型。因此，文章研究为分心驾驶行为检测领域 提供了一种更高效、更精确的解决方案，为道路安全和交通监管提供有力的支持。

关键词: 深度学习；注意力机制；驾驶员行为识别；目标检测算法