重庆邮电大学学报（自然科学版）2025，Vol.37Issue(1)：110-120,11.DOI:10.3979/j.issn.1673-825X.202405310133
动态运动约束下的视频异常检测
Dynamic motion constraints for video anomaly detection
摘要
Abstract
The low accuracy of existing frame-prediction-based methods for video anomaly detection is often attributed to the learning of non-compact normal distributions,leading to strong generalization over anomalous events.To address this issue,a framework named dynamic motion constraints for video anomaly detection(DMC-VAD)is proposed.A spatiotemporal constraint module is introduced,enabling the model to apply spatiotemporal constraints on normal motion based on optical flow salient objects and motion information.A spatiotemporal alignment and fusion module is designed to suppress low-quali-ty feature responses in optical flow using a correction alignment mechanism,enhancing the global appearance and motion context constraints of normal frames.These dynamic constraints allow the model to learn a more compact normal distribu-tion,with anomalous frames deviating from these constraints and thus producing high errors.Optical flow reconstruction is incorporated to better train the motion encoder,and a combination of optical flow mean error and peak signal-to-noise ratio of video frames is used as the anomaly score.On the UCSD Ped2 and ShanghaiTech datasets,the proposed model achieves the best area under the curve(AUC)for the receiver operating characteristic(ROC):99.42% (+1.19% )and 74.50% (+0.28% ),respectively.On the CUHK Avenue dataset,it achieves comparable results with AMSTE(88.07% )while re-ducing the parameter count to only 16% of AMSTE.关键词
视频异常检测/帧预测/无监督学习/动态约束/时空注意力Key words
video anomaly detection/frame prediction/unsupervised learning/dynamic constraint/spatiotemporal attention分类
信息技术与安全科学引用本文复制引用
石瑞年,何强,王恒友,陈琳琳..动态运动约束下的视频异常检测[J].重庆邮电大学学报（自然科学版）,2025,37(1):110-120,11.基金项目
国家自然科学基金项目(62072024,12301581) （62072024,12301581）
北京市教育委员会科学研究计划项目(KM202210016002,KM202110016001)National Natural Science Foundation of China(62072024,12301581) （KM202210016002,KM202110016001）
R&D Program of Beijing Municipal Educa-tion Commission(KM202210016002,KM202110016001) （KM202210016002,KM202110016001）
