Abstract: With the rapid advancement of intelligent driving technologies, the driver monitoring system (DMS) serves as a core module for road traffic safety guarantee and attracts growing industry attention. This paper constructs a driver monitoring system algorithm based on convolutional neural network (CNN) to improve driving safety and intelligence of commercial vehicles. The system consists of two major modules: cockpit monitoring and facial monitoring. The cockpit monitoring module adopts the YOLOv5 network structure. It learns driver behaviors such as sitting postures and hand movements through labeled data to identify hazardous driving behaviors including phone use and leaving the driving position. The facial monitoring module also applies the YOLOv5 network structure for training on self-built datasets. It automatically recognizes drivers' facial expressions and head postures to assess drivers' fatigue levels and attention status. Meanwhile, the system builds a classification detection framework based on MobileNet networks to distinguish camera occluded and non-occluded states. The YOLOv5n convolutional neural network integrated with Ghost modules cuts system parameters by 85%. It balances model lightweight design and feature recognition efficiency, maintains stable recognition accuracy under various working conditions, delivers favorable accuracy and real-time performance, and enhances the active safety monitoring capacity for drivers.

Key words: driver monitoring system; convolutional neural networks; YOLOv5; MobileNet

摘要： 随着智能驾驶技术的迅速发展，驾驶员监控系统（DMS）是保障道路交通安全的核心 模块，行业关注度持续提升。文章搭建基于卷积神经网络（CNN）的驾驶员监控系统算法， 提升商用车驾驶安全性与智能化水平，系统划分驾驶舱监测、面部监测两大模块。驾驶舱监 测模块依托 YOLOv5 网络结构，依托标注学习驾驶员坐姿、手部动作等行为，判别玩手机、 离岗等驾驶危险行为。面部监测模块沿用 YOLOv5 网络结构训练自建数据集，自主识别驾驶 员面部表情、头部姿态，判定驾驶员疲劳程度与注意力状态。同时，系统搭载 MobileNet 网 络搭建分类检测体系，区分摄像头遮挡与非遮挡状态。融合 Ghost 模块的 YOLOv5n 卷积神经 网络结构，可降低 85%系统参数量，兼顾模型轻量化与特征识别效率，适配多类工况且识别 精度稳定，具备良好准确性与实时性，可强化驾驶员主动安全监控效能。

关键词: 驾驶员监控系统；卷积神经网络；YOLOv5；MobileNet