燕山大学学报2025，Vol.49Issue(1)：74-82,9.DOI:10.3969/j.issn.1007-791X.2025.01.008
基于多尺度时空特征融合的视频异常事件检测
Anomaly detection in video based on multi-scale spatio-temporal features fusion
摘要
Abstract
In the problem of video anomalous event detection,existing research methods do not fully consider the background information interference and target scale variation in the scene,resulting in generally low detection accuracy.To address such problems,an anomalous event detection method that incorporates multi-scale spatio-temporal information is proposed.First,a coordinate attention approach is applied to make the model focus more on the regions where anomalous events occur.Second,to extract the rich spatio-temporal information at each level,a multi-branch and multi-scale feature fusion module is constructed using a dilated convolutional network.Finally,considering the diversity of normal behaviors,a regular score is proposed to further update the memory items in the memory-augmented module to improve the detection accuracy of anomalous events during the testing phase.In the experiments related to CUHK Avenue and ShanghaiTech datasets,the frame-level AUC of the proposed method reaches 88.7% and 77.5%,respectively,and meets the real-time requirements of video detection,which verifies the feasibility and effectiveness of the method.关键词
视频异常检测/无监督学习/空洞卷积/多尺度时空特征融合/记忆增强模块Key words
video anomaly detection/unsupervised learning/dilated convolution/multi-scaled spatio-temporal features fusion/memory-augmented module分类
计算机与自动化引用本文复制引用
李歌,肖洪兵,闫善武,王瑜,孙梅..基于多尺度时空特征融合的视频异常事件检测[J].燕山大学学报,2025,49(1):74-82,9.基金项目
北京市教育委员会科技计划重点项目(KZ202110011015) （KZ202110011015）
