SlowFast架构下景区异常行为识别算法及预警研究  

Research on abnormal behavior recognition algorithm and early warning in scenic spots under SlowFast architecture

作　　者：王志明[1] 张佳 彭江南 刘心志 陈克克 傅冠夷蛮 王绍萌 商飞[1] 狄长安[1] Wang Zhiming;Zhang Jia;Peng Jiangnan;Liu Xinzhi;Chen Keke;Fu Guanyiman;Wang Shaomeng;Shang Fei;Di Chang’an(School of Mechanical Engineering,Nanjing University of Science and Technology,Nanjing 210094,China;School of Computer Science and Engineering,Nanjing University of Science and Technology,Nanjing 210094,China)

机构地区：[1]南京理工大学机械工程学院,江苏南京210094 [2]南京理工大学计算机科学与工程学院,江苏南京210094

出　　处：《南京理工大学学报》2024年第3期374-383,共10页Journal of Nanjing University of Science and Technology

基　　金：国家重点研发计划(2021YFC1523500)。

摘　　要：针对当前古建筑场景下人员异常行为识别相关实例缺乏、数据集少、古建筑级别低、质量次,导致人员异常行为识别准确率低等问题,该文在古建筑景区背景下自行拍摄了多组视频,从中挑选构建了人员异常行为动作的5033段视频数据集:具有明确的典型古建筑背景;具有多人场景下暴恐打架斗殴、刻划、刻画以及存在火灾风险的人员异常行为等特征,并对每个视频进行了注释。该文首次于SlowFast网络框架中成功引入信号时域特征活动性、移动性参数,对构建的数据集进行高阶时序特征建模、增加分类算子。在人员异常行为识别任务中,模型的Top1准确率达到93.54%,而平均准确率达到96.30%,在SlowFast模型中引入活动性、移动性算子后,模型识别的准确率提升了0.87%。与几种常见架构的算法相比,该文所提出的方法存在一定的优势。Due to the lack of relevant examples,limited datasets,low level of ancient architectures,and poor quality of abnormal behavior recognition in the current scene of ancient architecture,the accuracy of abnormal behavior recognition is low.This paper independently shot multiple sets of videos in the background of ancient architectural scenic spots,and selected 5033 video datasets of personnel abnormal behavior from them:with clear typical ancient architectural backgrounds;with characteristics of violent terrorism fights,scratch,description,and abnormal behavior of individuals with fire risks in multi-person scenario,which each video were annotated.This paper successfully introduced signal temporal feature activity and mobility parameters into the SlowFast network framework for the first time,and modeled high-order temporal features and added classification operators to the constructed dataset.In the task of abnormal behavior recognition,the Top1 accuracy of the model reached 93.54%,while the mean accuracy reached 96.30%.After introducing activity and mobility operators into SlowFast model,the accuracy of model recognition improved by 0.87%.Compared with several common architecture algorithms,the method proposed in this paper has certain advantages.

关 键 词：机器视觉 异常行为识别 SlowFast 活动性算子 移动性算子 网格化预警 

分 类 号：TP389.1[自动化与计算机技术—计算机系统结构]