基于目标检测的驾驶人分神驾驶行为检测方法研究  

Study on Driver’s Distracted Driving Behavior Detection Method Based on Object Detection

作　　者：何廷全 俞山川 张生鹏 兰栋超 李刚[4] HE Ting-quan;YU Shan-chuan;ZHANG Sheng-peng;LAN Dong-chao;LI Gang(Guangxi New Development Transportation Group Co.,Ltd.,Nanning Guangxi 530029,China;Merchants Chongqing Communications Research&Design Institute Co.Ld.,Chongqing 400067,China;School of Electronic and Control Engineering,Chang'an University,Xi'an Shaanxi 710064,China;School of Energy and Electrical Engineering,Chang'an University,Xi'an Shaanxi 710064,China)

机构地区：[1]广西新发展交通集团有限公司,广西南宁530029 [2]招商局重庆交通科研设计院有限公司,重庆400067 [3]长安大学电子与控制工程学院,陕西西安710064 [4]长安大学能源与电气工程学院,陕西西安710064

出　　处：《公路交通科技》2022年第10期153-161,共9页Journal of Highway and Transportation Research and Development

基　　金：广西重点研发计划项目(桂科AB20159032);陕西省重点研发计划项目(2020ZDLGY09-03);国家山区公路工程技术研究中心开放基金项目(GSGZJ-2020-08)。

摘　　要：驾驶人的分神驾驶行为具有频发性和短暂性,极易引发道路交通安全事故,实时监测驾驶行为并对驾驶员的分神驾驶行为及时预警可有效降低车辆碰撞风险。针对该问题提出了一种基于深度学习和图像处理算法的驾驶人分神驾驶行为实时检测方法,对几种常见分神驾驶行为进行分类和检测。利用图像处理手段,分析分神驾驶行为检测的特点,在YOLOv5模型框架上进行了针对性的改进。首先,在网络中输入模块,采用动态优化方法设定了锚框信息,并在检测器头部之前添加特征选择操作,动态调整了特征点。针对主干网络,将BottlenckCSP网络的输入特征映射与输出特征直接连接,删除了模块的分支卷积,保留了更丰富的浅层特征信息。另一方面,在主干网络增加了注意力机制,提高网络模型提取关键特征的能力,从而提升模型的分类精度及鲁棒性。从总体上解决了特征信息丢失的问题,使模型易于完整的获取驾驶人的所有分神驾驶动作信息,从而减少深度学习网络参数,有效降低模型计算量。对Kaggle危险驾驶行为数据进行再标注,作为训练数据集,对模型进行了优化训练。结果表明:模型在测试集的平均检测精度为95.30%,平均召回率为95.13%,在试验环境下的检测速度达到48.3 FPS,满足分神驾驶实时监测的需求。The driver’s distracted driving behavior is frequent and transient,which is easy to cause road traffic accidents.Real time monitoring of driving behavior and timely warning of driver’s distracted driving behavior can effectively reduce the risk of vehicle collision.In view of this problem,a method for detecting driver’s distracted driving behavior based on deep learning and image processing algorithm is proposed to classify and detect several common driving behaviors.By using the image processing method and analyzing the characteristics of the distracted driving behavior detection,the targeted improvement is made on the YOLOv5 model.First,the module is input into the network,the anchor frame information is set by using the dynamic optimization method,and the feature selection operation is added before the detector head to dynamically adjust the feature points.For the backbone network,the input feature mapping of BotlenckCSP network is directly connected with the output feature,and the branch convolution of the module is deleted,leaving more abundant shallow feature information.On the other hand,the attention mechanism is added to the backbone network to improve the ability of network model to extract key features,thus improving the classification accuracy and robustness of the model.It solved the problem of loss of feature information as a whole,making it easy for the model to obtain all the distracted driving action information of the driver,thereby reducing the network parameters of deep learning and effectively reducing the computational burden of model.The Kaggle dangerous driving behavior data are re-labeled as the training data set,and the model is trained and optimized.The result shows that the average detection accuracy of the model in the test set is 95.30%,the average recall rate is 95.13%,and the detection speed in the test environment reaches 48.3 FPS,which meets the needs of real-time monitoring of distracted driving.

关 键 词：交通安全 分神驾驶 目标检测 实时监测 注意力机制 

分 类 号：U471.1[机械工程—车辆工程]