基于自编码器的人群异常行为检测算法  

Crowd anomaly behavior detection algorithm based on autoencoder

作　　者：王玉 杨晓文[1,2,3] 孙福盛 况立群 韩慧妍[1,2,3] 张元 WANG Yu;YANG Xiao-wen;SUN Fu-Sheng;KUANG Li-qun;HAN Hui-Yan;ZHANG Yuan(School of Computer Science and Technology,North University of China,Taiyuan 030051,China;Shanxi Provincial Key Laboratory of Machine Vision and Virtual Reality,Taiyuan 030051,China;Shanxi Province’s Vision Information Processing and Intelligent Robot Engineering Research Center,Taiyuan 030051,China)

机构地区：[1]中北大学计算机科学与技术学院,山西太原030051 [2]机器视觉与虚拟现实山西省重点实验室,山西太原030051 [3]山西省视觉信息处理及智能机器人工程研究中心,山西太原030051

出　　处：《计算机工程与设计》2025年第3期779-787,共9页Computer Engineering and Design

基　　金：山西省科技重大专项计划“揭榜挂帅”基金项目(202201150401021);国家自然科学基金项目(62272426);山西省科技成果转化引导专项基金项目(202104021301055);山西省自然科学基金项目(202303021211153、202203021212138)。

摘　　要：为提高人群异常行为检测算法性能,以STEAL-Net为基础,提出一种融合全局时空特征的自编码器人群异常行为检测算法。在编码器进行特征提取时,将全局跨通道特征提取模块与三维卷积结合,减少全局上下文特征的缺失;将提取到的特征序列输入到全局时空信息增强模块,进一步对视频帧的全局时空特征进行有效提取;进入解码器对输入帧进行重构,利用重构误差大小对异常行为进行检测。该算法在公开数据集UCSD Ped1、UCSD Ped2和ShanghaiTech上与其它先进方法进行了AUC指标的比较,实验结果表明所提算法的有效性。To enhance the efficacy of the algorithm for detecting anomaly behaviors in the crowd,an autoencoder algorithm based on STEAL-Net was proposed,in which global spatial-temporal features were integrated.In the encoder,the global cross-channel feature extraction module was combined with three-dimensional convolution,reducing the loss of global contextual features.The extracted feature sequences were input into the global spatio-temporal information enhancement module for further effective extraction of global spatio-temporal features of video frames.The input frames were reconstructed in the decoder,and the magnitude of reconstruction error was utilized for anomaly behavior detection.The proposed algorithm was compared with other advanced methods on public datasets UCSD Ped1,UCSD Ped2,and ShanghaiTech,based on the AUC metric.The effectiveness of the proposed algorithm is evidenced by the experimental results.

关 键 词：人群异常行为检测 自编码器 全局上下文 全局时空特征 重构 全局跨通道特征提取模块 全局时空信息增强模块 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]