摘要
无人机摄像头相较于固定摄像头具有更高的灵活度和适应性,将无人机与视频异常检测技术相结合具有重要的研究意义.主流的基于重建的异常检测方法都在同源数据中进行训练和测试,而对于数据分布不同的目标域场景下检测性能衰退.针对跨域的异常检测问题,设计了一个基于特征解耦的变形估计异常检测框架,该框架首先对源域和目标域训练样本进行特征解耦,提取分布对齐的域特征,将其馈送给多尺度变形模块和自编码网络.然后设计一个紧凑记忆模块,学习压缩的正常原型表示,并通过解码器对其重建.同时构建一个多尺度变形模块,利用K尺度的变形场估计图像的变形信息.最后对多尺度变形信息与重建误差加权求和生成异常分数.在4个视频数据集上进行域内和跨域实验,实验结果验证了提出的方法在视频异常检测方面的有效性.
Unmanned aerial vehicle(UAV)cameras offer greater flexibility and adaptability compared to fixed cameras,making the combination of UAVs with video anomaly detection technology significant for research purposes.Mainstream reconstruction-based anomaly detection methods are trained and tested within homogeneous data,but their detection performance declines in target domain scenarios with different data distributions.To address the cross-domain anomaly detection problem,a feature decoupling-based multi-scale deformation anomaly detection framework has been designed.This framework first performs feature decoupling on training samples from both the source and target domains to extract domain features that are aligned in distribution,which are then fed into a deformation estimated module and an autoencoder network.A compact memory module is designed to learn compressed representations of normal prototypes and reconstruct them through a decoder.Additionally,a multi-scale deformation module is constructed to estimate image deformation information using K-scale deformation fields.Finally,an anomaly score is generated by combining multi-scale deformation information with reconstruction errors through a weighted summation.Experiments are conducted within and across domains on four video datasets,and the results validate the effectiveness of the proposed method in video anomaly detection.
作者
穆辉宇
亢卿辰
谢毅
乔保军
远程
党兰学
MU Huiyu;KANG Qingchen;XIE Yi;QIAO Baojun;YUAN Cheng;DANG Lanxue(School of Computer and Information Engineering,Henan University,Henan Kaifeng 475004,China;Henan Key Laboratory of Big Data Analysis and Processing,Henan University,Henan Kaifeng 475004,China;Henan Provincial Fair Competition Review Affairs Center,Zhengzhou 467002,China)
出处
《河南大学学报(自然科学版)》
北大核心
2025年第1期12-20,共9页
Journal of Henan University:Natural Science
基金
2023年河南省研究生教育改革与质量提升工程项目(YJS2023JD28)
2024年河南省研究生教育改革与质量提升工程项目(YJS2024JD30)
河南省高校科技创新团队支持计划(24IRTSTHN021)
河南省科技攻关项目(232102210080,242102210079)
河南省高校重点科研项目(24A520004)
关键词
视频异常检测
自编码器
多尺度变形
记忆网络
域适应
video anomaly detection
autoencoder
multi-scale deformation
memory network
domain adaptation
