面向摄像头视频监控的泥石流发生场景智能识别方法  

Intelligent Identification Method of Debris Flow Scene Based on Camera Video Surveillance

作　　者：胡美辰 刘敦龙 桑学佳 张少杰[3] 陈乔[4] HU Mei-chen;LIU Dun-long;SANG Xue-jia;ZHANG Shao-jie;CHEN Qiao(College of Software Engineering,Chengdu University of Information and Technology,Chengdu 610225,China;Sichuan Province Informatization Application Support Software Engineering Technology Research Center,Chengdu 610225,China;Institute of Mountain Hazards and Environment,Chinese Academy of Sciences,Chengdu 610041,China;Chongqing Institute of Green Intelligent Technology,Chinese Academy of Sciences,Chongqing 400714,China)

机构地区：[1]成都信息工程大学软件工程学院,四川成都610225 [2]四川省信息化应用支撑软件工程技术研究中心,四川成都610225 [3]中国科学院水利部成都山地灾害与环境研究所,四川成都610041 [4]中国科学院重庆绿色智能技术研究院,重庆400714

出　　处：《计算机与现代化》2024年第3期41-46,共6页Computer and Modernization

基　　金：国家自然科学基金青年项目(42001100);四川省自然科学基金资助项目(2023NSFSC0751);四川省信息化应用支撑软件工程技术研究中心开放课题(760115027)。

摘　　要：摄像头视频监控在泥石流防灾减灾中的应用较为广泛,但现有的视频检测技术功能有限,无法自动判断出泥石流灾害事件的发生。针对这一问题,本文基于迁移学习策略,改进一种基于卷积神经网络的视频分类方法。首先,借助TSN模型框架,将底层网络架构更改为ResNet-50,用于运动特征提取和泥石流场景识别。然后,通过ImageNet和Kinet-ics-400数据集预训练该模型,使模型具备较强的泛化能力。最后,结合经过预处理的地质灾害视频数据集对模型进行训练和微调,使其能够精准地识别出泥石流事件。通过大量的运动场景视频对该模型进行检验,实验结果表明,该方法对泥石流运动场景视频的识别准确率可达87.73%。因此,本文的研究成果可充分发挥视频监控在泥石流监测预警中的作用。Camera video surveillance is widely used in debris flow disaster prevention and mitigation,but the existing video de⁃tection technology has limited functions and can not automatically judge the occurrence of debris flow disaster events.To solve this problem,using transfer learning strategy,this paper improves a video classification method based on convolutional neural network.Firstly,with the help of TSN model framework,the underlying network architecture is changed to ResNet-50,which is utilized for motion feature extraction and debris flow scene identification.Then,the model is pre-trained with ImageNet and Ki⁃netics 400 datasets to make the model have strong generalization ability.Finally,the model is trained and fine-tuned with the pre-processed geological disaster video dataset,so that it can accurately identify debris flow events.The model is tested by a large number of moving scene videos,and the experimental results show that the identification accuracy of the method for debris flow movement video can reach 87.73%.Therefore,the research results of this paper can to the play a full role of video surveil⁃lance in debris flow monitoring and warning.

关 键 词：泥石流 视频监控 运动场景 迁移学习 智能识别 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]