煤矿视频中复杂行为识别的持续学习模型探究  

Research on Continuous Learning Model of Complex Behavior Recognition in Coal Mine Video

作　　者：罗响 袁艳斌[1] 王德永[1,2] 钟珊 张波 李倩[1] LUO Xiang;YUAN Yanbin;WANG Deyong;ZHONG Shan;ZHANG Bo;LI Qian(School of Resources and Environmental Engineering,Wuhan University of Technology,Wuhan 430070,China;School of Information Engineering,Pingdingshan University,Pingdinshan 467000,China)

机构地区：[1]武汉理工大学资源与环境工程学院,湖北武汉430070 [2]平顶山学院信息工程学院,河南平顶山467000

出　　处：《金属矿山》2020年第10期118-123,共6页Metal Mine

摘　　要：为更好保障矿工井下作业安全,如何提高矿井监控视频中矿工复杂行为识别准确率已成为研究热点。通过耦合深度网络和主动学习方法构建的矿工复杂行为持续学习模型,可自动对新增样例进行标记,并持续从视频数据中学习人体行为,从而提高识别准确率。分析在是否为主动学习和是否固定缓冲区大小4种实验环境下公共数据集KTH和真实煤矿监控数据集RCV中复杂行为的识别性能,发现随着新增样例的加入,持续学习模型框架能够不断改进每种复杂行为模型的识别性能,且最终的识别准确率相较于传统识别模型有明显提升。结果表明持续学习行为模型能有效解决复杂行为识别过程中的概念漂移问题,且具有良好的自学习能力和鲁棒性。To better ensure the safety of miners’ underground operations,how to improve the recognition accuracy ofminers’ complex behaviors in mine surveillance videos has become a research hotspot. The continuous learning model of com-plex behaviors of miners constructed by coupling deep networks and active learning methods can automatically mark new ex-amples and continuously learn human behaviors from video data,thereby improving recognition accuracy. This paper analyz-es the recognition performance of complex behaviors in the public data set KTH and the real coal mine monitoring data setRCV in the four experimental environments of whether it is active learning and whether the buffer size is fixed,it is foundthat with the addition of new examples,the continuous learning model framework can continuously improve the recognitionperformance of each complex behavior model,and the final recognition accuracy is significantly improved compared to tradi-tional recognition models. The results show that the continuous learning behavior model can effectively solve the problem ofconcept drift in the process of complex behavior recognition and has good self-learning ability and robustness.

关 键 词：煤矿工人 视频监控 复杂行为识别 持续学习 

分 类 号：TF841[冶金工程—有色金属冶金] TD983[矿业工程—选矿]