基于视频分析的空管员违规行为识别方法  

Violations recognition method of air traffic controllers based on video analysis

作　　者：唐豪 奉鑫鑫 高曙[1] 罗帆[2] 揣明瑞 TANG Hao;FENG Xinxin;GAO Shu;LUO Fan;CHUAI Mingrui(School of Computer Science&Artificial Intelligence,Wuhan University of Technology,Wuhan Hubei 430070,China;School of Management,Wuhan University of Technology,Wuhan Hubei 430070,China)

机构地区：[1]武汉理工大学计算机与人工智能学院,湖北武汉430070 [2]武汉理工大学管理学院,湖北武汉430070

出　　处：《中国安全生产科学技术》2023年第1期196-201,共6页Journal of Safety Science and Technology

基　　金：国家自然科学基金项目(71271163);国家教育部人文社科项目(18YJA630076)。

摘　　要：为利用视频数据对空管员违规行为进行智能化分析,降低不安全事件发生率,提出2阶段的违规行为识别模型(AR-ResNeXt),基于实地调研构建空管员视频数据集,利用最小化动态多实例学习损失函数和中心损失函数,获得违规行为检测的判别特征表示,结合异常回归网络和ResNeXt网络,完成对空管员违规行为的时序区间检测与动作分类。研究结果表明:AR-ResNeXt模型在自制数据集中,其帧级AUC达到82.9%,分类准确率达到87.8%,可准确识别空管员发生违规行为的时序区间并进行分类,研究结果可为保障空中交通安全奠定基础。In order to use video data to intelligently analyze the violations of air traffic controllers and reduce the incidence of unsafe events,a two-stage violations recognition model(AR-ResNeXt)was proposed.Based on the video data set of air traffic controllers constructed by field investigation,the discriminant feature representation of violations detection was obtained by minimizing the dynamic multiple-instance learning loss function and the center loss function,and the time ordered interval detection and action classification of violations of air traffic controllers were completed by combining with the abnormal regression network and ResNext network.The results showed that the AR-ResNeXt model achieved a frame-level AUC of 86.50%and a classification accuracy of 87.8%on the self-made data set.It can accurately identify and classify the time ordered interval of violations of air traffic controllers,and lay a foundation for ensuring the air traffic safety.

关 键 词：违规行为 深度学习 视频监控 空中交通管制员 异常回归网络 ResNeXt网络 

分 类 号：X949[环境科学与工程—安全科学]