摘要
因传统监控领域存在监控成本高、需要大量人力投入等问题,普通监控模式已经难以满足社会发展需要,发展智慧监控迫不容缓。针对上述问题,提出一种基于深度学习的打架行为监测预警原型系统。该系统通过对开源人体姿态识别项目OpenPose进行简化改进,在确保一定准确率的情况下,精简网络结构,减少计算参数,进而压缩模型大小。同时配合人体打架行为识别人工规则,对二维视频中可能存在的打架行为进行预警,从而降低监控人员负担,提升监控效率,及时制止打架行为,避免暴力事件出现。实验结果表明,经简化过的系统输出模型准确度可达到原系统输出模型准确度70%以上,模型大小同比缩小50%以上,可以成功预警打架行为,并标明打架行为动作实施者,符合设计需求,达到实用标准。
Due to the problems of high monitoring cost and large amount of human input in the traditional monitoring field,ordinary monitoring models cannot meet the needs of social development,and the development of intelligent monitoring cannot be delayed. To solve the above problems,a prototype system of fighting behavior monitoring and early warning based on deep learning was proposed. The system simplified and improved the open-source human gesture recognition project OpenPose,and under the condition of ensuring a certain accuracy,the network structure was simplified,the calculation parameters were reduced,and the model size was reduced. At the same time,it cooperated with the manual rules of human body fighting behavior identification to provide early warning of possible fighting behaviors in the twodimensional video,thereby reducing the burden of monitoring personnel,improving monitoring efficiency,stopping fighting behaviors in time,and avoiding violent incidents. The experimental results show that the accuracy of the simplified system output model can reach more than 70% of that of the original system output model,and the model size is reduced by more than 50%. It can successfully warn the fighting behavior and indicate the implementer of the fighting behavior,which meets the design requirements.
作者
马子健
林雨衡
王志强
都迎迎
MA Zijian;LIN Yuheng;WANG Zhiqiang;DU Yingying(College of Computer and Information Technology,Beijing Jiaotong University,Beijing 100044,China;Department of Cyberspace Security,Beijing Electronic Science and Technology Institute,Beijing 100070,China)
出处
《计算机应用》
CSCD
北大核心
2021年第S02期214-220,共7页
journal of Computer Applications
基金
国家重点研发计划项目(2017YFC1201204)
中国博士后科学基金面上项目(2019M650606)
北京电子科技研究院一流学科建设项目(3201012)。
关键词
智慧监控
深度学习
人体姿态识别
行为监测
模型简化
smart monitoring
deep learning
human body posture recognition
behavior monitoring
model simplification