摘要
随着姿态估计技术的发展,使用人体骨骼数据而非传统像素数据进行异常行为检测成为可能,这种异常检测方法克服了传统基于像素特征的方法中光照、视角和背景噪声等因素带来的问题。然而,现有的图卷积网络(Graph Convolutional Network,GCN)在处理人体骨骼数据时,通常使用固定的邻接矩阵进行信息聚合,这限制了模型在提取行为特征时的灵活性。为了解决上述问题,本文提出了基于时间-通道拓扑解耦的图卷积网络(TimeChannel Topology Decoupling Graph Convolution Network,TCTD-GCN)。TCTD-GCN分别在时间和通道维度上采用拓扑学习的方式来学习自适应的邻接矩阵,从而实现时间和通道的有效解耦。学习得到的自适应邻接矩阵能更准确地聚合特征,促进对人体行为的准确表示。此外,文章提出一种虚拟异常引导的自监督异常检测(Virtual Anomaly-guided Self-supervised Anomaly Detection,VASAD)策略来提高检测精度。VASAD将异常检测问题视作一个多分类问题,通过将正常行为的不同类别视为“虚拟异常”来辅助模型训练,从而在测试阶段更准确地区分正常与异常行为。这种策略增强了模型对正常行为内在差异的学习,提高了对真实异常的判别能力。最后,本文模型在ShanghaiTech Campus、CUHK Avenue和USCD Ped2三个主流数据集上进行实验,帧级曲线下面积(area under the curve,AUC)分别达到76.6%、87.7%和95.3%,在ShanghaiTech Campus和CUHK Avenue数据集上相对主流模型有明显提升,验证了模型的有效性和优越性。
With the development of pose estimation technology,human skeleton data can be used instead of conven‐tional pixel data for abnormal behavior detection.This anomaly detection method overcomes the problems caused by illu‐mination,view angle,and background noise in conventional pixel-based methods.However,the existing graph convo‐lutional network(GCN)usually uses a fixed adjacency matrix for information aggregation when processing human skel‐eton data,which limits the flexibility of the model in extracting behavioral features.To solve this problem,this study proposes a time-channel topology decoupling graph convolution network(TCTD-GCN).TCTD-GCN adopts topologi‐cal learning to learn an adaptive adjacency matrix in the time and channel dimensions,respectively,so as to realize their effective decoupling.The learned adaptive adjacency matrix can aggregate features more precisely and promote the accu‐rate representation of human behavior.In addition,we propose a virtual anomaly-guided self-supervised anomaly detec‐tion(VASAD)strategy to improve the detection accuracy.VASAD treats the anomaly detection problem as a multi-classification problem,and treats different classes of normal behavior as“virtual anomalies”to assist model training,thus distinguishing normal and abnormal behavior more accurately in the testing phase.This strategy enhances the model’s learning of the inherent differences in normal behavior and improves the discrimination ability of true anoma‐lies.The model was tested on three mainstream datasets,ShanghaiTech Campus,CUHK Avenue,and USCD Ped2,and the frame-level area under the curve reached 76.6%,87.7%,and 95.3%,respectively.Compared with mainstream models,the proposed model significantly improved the detection performance on the ShanghaiTech Campus and CUHK Avenue datasets,which verifies its effectiveness and superiority.
作者
张家波
唐上松
何阿娟
ZHANG Jiabo;TANG Shangsong;HE Ajuan(School of Communication and Information Engineering,Chongqing University of Posts and Telecommunications,Chongqing 400065,China)
出处
《信号处理》
CSCD
北大核心
2024年第12期2193-2205,共13页
Journal of Signal Processing
基金
国家自然科学基金面上项目(62171070)
重庆市研究生科研创新项目(CYS23456)。
关键词
人体骨骼
异常行为检测
解耦图卷积
自适应邻接矩阵
多类别异常检测
human skeleton
abnormal behavior detection
decoupling graph convolution
adaptive adjacency ma‐trix
multi-category anomaly detection
