基于深度学习的驾驶员分心行为检测方法  

Driver distraction behavior detection method based on deep learning

作　　者：赵栓峰[1] 李乐平 王茂权 李小雨 谢乐坤 Zhao Shuanfeng;Li Leping;Wang Maoquan;Li Xiaoyu;Xie Lekun(College of Mechanical Engineering,Xi′an University of Science and Technology,Xi′an 710054,China)

机构地区：[1]西安科技大学机械工程学院,西安710054

出　　处：《电子测量技术》2024年第9期145-153,共9页Electronic Measurement Technology

基　　金：陕西省自然科学基础研究计划面上项目(2024JC-YBMS-366);陕西省重点研发计划(2020ZDLGY04-06)项目资助。

摘　　要：驾驶员分心行为检测对于开发以驾驶员为中心的人车协同驾驶系统具有至关重要的意义。针对现有基于卷积神经网络的驾驶员分心行为检测模型缺乏全局特征提取能力、泛化性能较弱以及忽视了驾驶场景中不同区域的重要性,构建一种基于深度学习的驾驶员分心行为检测模型,实现对驾驶员分心行为的准确检测。首先,开发了基于HorNet的残差结构,通过高阶空间交互来增强特征表示能力;其次,受人类注意力机制以及现有注意力机制的启发,设计一种自适应加权注意策略来提取与驾驶行为最相关的特征;然后,在现有的分类数据集上训练本文模型,并使用先验知识作为初始权值来改善训练结果,进而提高模型的泛化能力;最后,对驾驶行为特征进行可视化,以提高人们对于本文模型的信任度。实验结果表明,本文模型可以准确地检测驾驶员分心行为,在准确性和可靠性方面明显优于现有方法。Driver distraction behaviour detection is of crucial significance for the development of driver-centered human-vehicle co-driving systems.Aiming at the existing convolutional neural network-based driver distraction detection models that lack global feature extraction capability,have weak generalisation performance and neglect of the importance of different regions in the driving scene,a driver distraction detection model based on deep learning is constructed to achieve accurate prediction of driver distraction behaviour.First,a residual structure based on HorNet is developed to enhance the feature representation capability through higher-order spatial interactions;second,inspired by the human attention mechanism and the existing attention mechanisms,an adaptive weighted attention strategy is designed to extract the features most relevant to the driving behaviour;and then,the model in this paper is trained on the existing categorical dataset,and the a priori knowledge is used as the initial weights to improve the training results which in turn improves the generalisation ability of the model;finally,the driving behaviour features are visualised to improve the trust in this paper′s model.The experimental results show that the model in this paper can accurately detect driver distraction behaviour,which is significantly better than existing methods in terms of accuracy,and reliability.

关 键 词：驾驶员行为 先验知识 深度学习 视觉可解释性 

分 类 号：TN0[电子电信—物理电子学] TP391[自动化与计算机技术—计算机应用技术]