基于Yolo和GOTURN的景区游客翻越行为识别  

A Fence Climbing Behavior Recognition of Scenic Area Tourist Based on Yolo and GOTURN

作　　者：周巧瑜 曹扬 詹瑾瑜[1,2] 江维 李响[2,3] 杨瑞 ZHOU Qiao-yu;CAO Yang;ZHAN Jin-yu;JIANG Wei;LI Xiang;YANG Rui(School of Information and Software Engineering,University of Electronic Science and Technology of China,Chengdu 610054,China;CETC Big Data Research Institute Co.,Ltd.,Guiyang 550022,China;Big Data Application on Improving Government Governance Capabilities National Engineering Laboratory,Guiyang 550022,China)

机构地区：[1]电子科技大学信息与软件工程学院,四川成都610054 [2]中电科大数据研究院有限公司,贵州贵阳550022 [3]提升政府治理能力大数据应用技术国家工程实验室,贵州贵阳550022

出　　处：《计算机技术与发展》2022年第1期134-140,共7页Computer Technology and Development

基　　金：提升政府治理能力大数据应用技术国家工程实验室开放基金项目(W-2019007);四川省科技项目(2018CC0136);中科院计算机体系结构国家重点实验室开放课题(CARCH201811);中央高校基本科研业务费(ZYGX2018J077,ZYGX2019J078)。

摘　　要：近年来,随着旅游市场的快速发展,在旅游景区出现的一些违规行为,不仅危害了人身安全,而且也给社会造成了许多负面影响。由于出现该类行为的频率不高,通过人工观察耗费大量人力资源且效率不高,使用深度学习算法对具体行为进行识别,帮助景区监管人员快速预警违规行为,已成为必然趋势。针对这一问题,结合目标检测与目标跟踪任务,该文提出了一种基于Yolo和GOTURN的景区游客翻越行为识别方法。首先将视频转为视频帧,再经过Yolo目标检测和GOTURN目标跟踪得到人员边界框坐标和视频帧轨迹点集合,再进入轨迹分析得出最终结果标签(是否为翻越行为),形成一个完整的翻越行为识别方法。实验数据表明,基于Yolo和GOTURN的景区游客翻越行为识别方法相对于其他方法具有较高的准确率,应用在实际的景区游客翻越行为识别系统中得到了93.7%的准确率。In recent years,with the rapid development of tourism market,there are some tourism violation behaviors,which not only endanger personal safety,but also cause many negative effects on the society.Due to the infrequent occurrence of such behaviors,manual observation costs a lot of human resources and is inefficient.It has become an inevitable trend to use deep learning algorithms to identify specific behaviors and help scenic area supervisors to quickly warn violation behaviors.For this,combining target detection and target tracking tasks,we introduce a fence climbing behavior recognition method for the scenic area tourists based on Yolo and GOTURN.Firstly,the video is converted to video frame,and then the boundary frame coordinates and the video frame track point set are obtained by Yolo target detection and GOTURN target tracking.Finally,through trajectory analysis,the final result label(whether it is a fence climbing behavior or not)is obtained to form a fence climbing behavior recognition method.The experiment shows that the proposed fence climbing behavior recognition method has higher accuracy compared with the other methods,and the accuracy of 93.7% is obtained in the actual scenic spot tourist jump behavior recognition system.

关 键 词：深度学习 目标检测 目标跟踪 翻越行为识别 Yolo GOTURN 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]