华中科技大学学报（自然科学版）2025，Vol.53Issue(10)：8-14,7.DOI:10.13245/j.hust.251093
基于姿态-场景特征的视频异常检测研究
Research on video anomaly detection based on pose-scene features
摘要
Abstract
The low-dimensional and highly structured characteristics of pose skeleton points were utilized,and graph convolutional networks(GCNs)were employed to hierarchically and structurally process pose features,analyzing both individual poses and the interactions between individuals.The rich semantic information of the scene was incorporated,and the conditional variational autoencoder(CVAE)was introduced for anomaly detection.The CVAE encoded scene images and pose feature mappings,generating pose-scene conditional feature maps that enhanced the fusion of scene and pose features and improved anomaly detection accuracy.The proposed model effectively integrated pose and scene features,significantly improving the detection of abnormal behaviors in complex environments.On the Shanghai Tech,CUHK Avenue,and NWPU campus anomaly detection datasets,the proposed model achieves area under the curve(AUC)performances of 84.3％,87.2％,and 69.7％,respectively,demonstrating its superiority compared with existing methods.关键词
姿态估计/图卷积神经网络/条件变分自编码器/分层结构/视频异常检测Key words
pose estimation/graph convolutional neural network/conditional variational autoencoder/hierarchical structure/video anomaly detection分类
计算机与自动化引用本文复制引用
陈志刚,张心宇,刘凌枫,李航..基于姿态-场景特征的视频异常检测研究[J].华中科技大学学报（自然科学版）,2025,53(10):8-14,7.基金项目
科技创新2030——"新一代人工智能"重大项目(2020AAA0109605) （2020AAA0109605）
长沙市科技计划重大专项(kh2103016). （kh2103016）
