基于深度学习的矿井视频流异常检测算法研究  

Research on anomaly detection algorithm of video stream in mine based on deep learning

作　　者：索智文 丁剑明 屈波 张兰峰 申茂良 SUO Zhiwen;DING Jianming;QU Bo;ZHANG Lanfeng;SHEN Maoliang(Guoneng Shendong Coal Intelligent Technology Center,Yulin Shaanxi 719315,China;Shaanxi EKIA Information Technology Co.,Ltd.,Xi’an Shaanxi 710065,China;China Coal Research Institute Co.,Ltd.,Beijing 100013,China)

机构地区：[1]国能神东煤炭智能技术中心,陕西榆林719315 [2]陕西亿杰鑫信息技术有限公司,陕西西安710065 [3]煤炭科学技术研究院有限公司,北京100013

出　　处：《中国安全生产科学技术》2025年第3期133-140,共8页Journal of Safety Science and Technology

摘　　要：为了探究矿井复杂环境中视频流检测精度问题,提出1种基于YOLOv4深度优化的复杂环境视频流异常检测算法,增设SE模块提升特征提取效率,改进SPP、PANet模块优化异常检测能力;提取矿井现场真实数据,对数据集中4500多张异常行为进行模型训练,采用深度优化的YOLOv4算法进行识别,标注出视频异常行为。研究结果表明:相较于传统的YOLOv4算法,深度优化后的模型平均精确率均值(MAP)为98.02%,MAP提升16.6百分点,每秒传输帧数(FPS)提高至28.56。研究结果可为优化矿井复杂环境下视频流检测精度提供思路和方法。To investigate the issue of video stream detection accuracy in complex environment of mine,an anomaly detection algorithm of video stream in complex environment based on deeply optimized YOLOv4 was proposed.A SE module was added to enhance the feature extraction efficiency,and the SPP and PANet modules were improved to optimize the anomaly detection capability.The on-site real data of mine was extracted,and over 4500 abnormal behavior in the dataset were used for model training.The deeply optimized YOLOv4 algorithm was employed for the identification,and the abnormal behavior in the video was annotated.The results show that tcomparing with the traditional YOLOv4 algorithm,the mean Average Precision(MAP)of the deeply optimized model was 98.02%,increasing by 16.6 percentage points,and the Frames Per Second(FPS)of transmission was increased to 28.56.The research findings can provide insights and methods for optimizing the accuracy of video stream detection in complex environment of mines.

关 键 词：YOLOv4算法 视频监控 视频流异常检测 MAP 矿山智能化 

分 类 号：X936[环境科学与工程—安全科学]