双端驾乘人员车云协同计算危险行为分析  

Dual-end Analysis of Drivers′and Passengers′Dangerous Behaviors with Vehicle-cloud Collaborative Computing

作　　者：井晶[1,2] 贺晨 刘营 赵作鹏 JING Jing;HE Chen;LIU Ying;ZHAO Zuopeng(Xuzhou Finance and Economics Branch,Jiangsu United Vocational and Technical College,Xuzhou Jiangsu 221008,China;School of Computer Science and Technology,China University of Mining and Technology,Xuzhou Jiangsu 221116,China)

机构地区：[1]江苏联合职业技术学院徐州财经分院,江苏徐州221008 [2]中国矿业大学计算机科学与技术学院,江苏徐州221116

出　　处：《现代雷达》2024年第11期110-117,共8页Modern Radar

基　　金：国家自然科学基金资助项目(61976217)。

摘　　要：车内人员不安全行为的预警是减少交通事故的重要手段,然而现有的研究主要关注于驾驶员的危险行为检测,乘客的危险行为往往被忽视,因此文中提出了一种基于车云算力协同的驾乘人员危险行为分析算法。首先,通过基于粗粒度深度估计的检测和人员身份分析模型区分驾驶员与乘客;然后,通过基于ByteTrack多目标跟踪的检测模型检测乘客是否越位,并通过头部姿态估计与疲劳状态检测的多任务联合模型检测驾驶员的危险驾驶行为;最后,将危险行为视频片段上传云端进行校验。实验结果表明,文中提出的基于粗粒度深度估计的检测模型和人员身份分析模型相比感兴趣区域法的F_(1)-Score提高了5.1%,多任务联合模型中的头部姿态估计分支相比LwPoser模型F_(1)-Score提高了5.2%,疲劳检测分支相比YOLOv6s模型平均准确率均值提高了2.4%,且校验中使用云端高精度模型F_(1)-Score提高了7.6%。以上结果证明了文中所提算法的有效性。The warning of unsafe behavior inside the vehicle is an important means to reduce traffic accidents.However,existing research mainly focuses on detecting dangerous behaviors of drivers,while dangerous behaviors of passengers are often overlooked.Therefore,an algorithm for analyzing dangerous behaviors of both drivers and passengers based on vehicle-cloud computing collabo-ration is proposed in this paper.First,the driver and passengers are distin guished using a detection and personnel identity analysis model based on coarse-grained depth estimation.Then,a detection model based on ByteTrack multi-object tracking is used to de-tect whether passengers are out of their seats,and a multi-task joint model for head pose estimation and fatigue state detection is employed to detect dangerous driving behaviors of the driver.Finally,video clips of dangerous behaviors are uploaded to the cloud for verification.Experimental results show that the proposed detection model based on coarse-grained depth estimation and the per-sonnel identity analysis model increase F_(1)-Score by 5.1%compared to the region of interest method,the head pose estimation branch of the multi-task joint model increases F_(1)-Score by 5.2%compared to the LwPoser model,the fatigue detection branch in-creases the mean average precision by 2.4%compared to the YOLOv6s model,and the use of high-precision cloud-based models increases F_(1)-Score by 7.6%during venification.The above results demonstrate the flectiveness of the proposed algorithm.

关 键 词：车云算力协同 深度估计 多任务联合 目标检测 目标跟踪 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]