基于显著特征和时空图网络的视频异常事件检测  

Video abnormal event detection based on salient features and space-time fusion graph network

作　　者：谭奕鑫 詹永照[1] 刘洪麟 TAN Yixin;ZHAN Yongzhao;LIU Honglin(School of Computer Science and Communication Engineering,Jiangsu University,Zhenjiang,Jiangsu 212013,China)

机构地区：[1]江苏大学计算机科学与通信工程学院,江苏镇江212013

出　　处：《江苏大学学报(自然科学版)》2025年第2期179-188,共10页Journal of Jiangsu University:Natural Science Edition

基　　金：国家自然科学基金资助项目(61672268)。

摘　　要：针对目前基于片段的视频异常事件检测模型仅考虑片段的正常或异常会产生异常碎片化和异常起止位置蔓延以及异常事件分类难的问题,提出一种基于显著特征和时空图网络的视频异常事件检测与分类方法.首先,提出基于时空融合图网络的视频异常时序片段整合与精化方法,整合出连续的异常区域,同时考虑时空融合图网络的特征传递性,精化异常区域,以有效解决异常判别存在不确定性和异常片段碎片化问题.其次,针对弱监督异常事件内在特征难以有效表达而引起分类难的问题,提出异常事件特征学习和分类方法,在异常区域中建立特征相似图和异常相似图,并利用图卷积网络融合学习异常事件的特征,设计类别不平衡损失函数,从而提高异常事件的分类性能.在UCF-Crime数据集中进行试验,结果表明:文中方法的曲线下面积AUC达到了85.37%,比基准线SULTANI方法高9.83%,比最好的同类方法THAKARE方法高0.89%;在异常事件分类上,该方法获得了74.06%的平均准确率,比现有ZHOU方法提高4.39%.该方法能更有效检测定位视频异常事件,且异常事件分类性能更优.To solve the problems that the current segment-based video abnormal event detection model with only considering the normal or abnormal of segments could lead to the abnormal fragmentation,the abnormal start/stop location spreading and the difficulty of abnormal event classification,the video abnormal event detection and classification method was proposed based on salient features and space-time graph network.The video anomaly temporal fragment integration and refinement method based on space-time fusion graph network was proposed to integrate the continuous anomaly regions.Considering the feature transferability of the space-time fusion graph network,the anomaly regions were refined to effectively solve the problems of uncertainty in anomaly discrimination and fragmentation of anomaly fragments.To solve the difficult classification problem of weakly supervised anomalous events caused by the intrinsic features of difficultly effective expressing,the abnormal event feature learning and classification method was proposed.The feature similarity graphs and abnormal similarity graphs were established in abnormal regions,and the graph convolutional networks were used for the abnormal event feature fusion learning.The abnormal-imbalance loss function was designed to improve the classification performance of abnormal events.The results show that by the proposed method,the AUC reaches 85.37% on the UCF-Crime dataset,which is 9.83% higher than the baseline of SULTANI method and 0.89% higher than that by the state-of-the-art method of THAKARE method.For the abnormal event classification,the average accuracy of 74.06% is achieved,which is increased by 4.39% compared with the existing ZHOU method.The proposed method can detect and locate video anomalous events more effectively with better anomalous event classification performance.

关 键 词：视频异常检测 异常事件分类 特征相似图 显著特征序列 异常事件特征学习 类别不平衡损失函数 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]