基于图像和深度学习的驾驶人状态监测算法综述  

Review of Image and Deep Learning Based Algorithms for Driver State Monitoring

作　　者：王畅[1] 李朝 赵霞 孙秦豫 付锐[1] 郭应时[1] 袁伟[1] WANG Chang;LI Zhao;ZHAO Xia;SUN Qin-yu;FU Rui;GUO Ying-shi;YUAN Wei(School of Automobile,Chang'an University,Xi'an 710018,Shaanxi,China;Automotive Engineering Research Institute,Jiangsu University,Zhenjiang 212013,Jiangsu,China)

机构地区：[1]长安大学汽车学院,陕西西安710018 [2]江苏大学汽车工程研究院,江苏镇江212013

出　　处：《中国公路学报》2025年第1期324-347,共24页China Journal of Highway and Transport

基　　金：国家自然科学基金项目(52272412,52362050);江苏省自然科学基金项目(BK20240870);长安大学中央高校基本科研业务费专项资金项目(300102224501,300102224302)。

摘　　要：驾驶人状态监测技术作为提升车辆智能化水平与安全性的关键手段,旨在精确辨识与深入理解驾驶人的动作、情绪及注意力状态。尽管该领域研究已取得显著进展,但针对深度学习算法原理的系统性总结尚显不足。鉴于此,系统性地综述了基于图像与深度学习的驾驶人状态监测算法,以满足智能车辆技术持续发展的需求。首先阐述了文献综述方法;接着对现有公开数据集进行了整理与描述;然后从数据选择与处理、模型架构、模型训练与评估以及优化目标几个方面分别展开了深入阐述;最后总结了目前研究中存在的不足,并对未来研究的主要发展方向进行了展望。结果表明:基于图像和深度学习的驾驶人状态监测研究已进展到一定深度,数据选择与处理技术呈现多样性,模型架构朝着多模态、多任务、轻量化和高鲁棒性的方向不断发展,并逐渐开始采用非完全监督和多目标优化的训练策略。然而,大多数研究方法缺乏针对实际驾驶场景的系统性测试,也未充分考虑自然驾驶条件下驾驶人行为特性以及智能车辆人机交互形态的变化,因此难以实现对各种驾驶场景和驾驶人个性的全方位监测功能。驾驶人状态监测算法的进一步发展主要受限于两方面因素。其一,目前深度学习方法在其领域适应性、可解释性、运行效率等方面仍存在不足;其二,缺乏自然驾驶环境下大规模高质量数据集。该综述可为高认知驾驶人状态监测系统的进一步发展提供有效指导和重要参考。Driver state monitoring technology,as a key means for improving vehicle intelligence and safety,aims to accurately identify and deeply understand the driver's actions,emotions,and attention states.Although significant progress has been made in this field,a systematic summary of the principles of deep learning algorithms is lacking.In view of this,this paper systematically reviews driver state monitoring algorithms based on images and deep learning to meet the needs of the continuous development of intelligent vehicle technology.First,the methodology in the literature is elaborated upon.The existing publicly available datasets are then organized and described.Subsequently,in-depth exploration is conducted from the aspects of data selection and processing,model architecture,model training and evaluation,and optimization.Finally,the shortcomings of the current research are summarized,and the main future development directions are outlined.The results show that:①the research on driver state monitoring based on image and deep learning has progressed to a certain depth;②data selection and processing techniques show diversity;③model architectures continue to evolve in the direction of multi-modal,multitasking,lightweight,and high robustness,gradually beginning to adopt training strategies for incomplete supervision and multi-objective optimization.However,most research methods lack systematic testing of actual driving scenarios neither fully considering the behavioral characteristics of drivers under natural driving conditions nor the changes in the human-computer interaction patterns of intelligent vehicles,making it difficult to construct an all-around monitoring function for various driving scenarios and driver personalities.The further development of driver state monitoring algorithms is mainly limited by two factors.First,the current deep learning methods still have deficiencies in their domain adaptation,interpretability,and operational efficiency.Second,large-scale high-quality datasets under natural drivin

关 键 词：汽车工程 智能驾驶 综述 驾驶人状态监测 深度学习 计算机视觉 模型架构 训练策略 

分 类 号：U471.3[机械工程—车辆工程]