基于深度学习的前沿视频异常检测方法综述  

Frontier video anomaly detection methods based ondeep learning:comprehensive review

作　　者：李南君 聂秀山 李拓 邹晓峰 王长红 Li Nanjun;Nie Xiushan;Li Tuo;Zou Xiaofeng;Wang Changhong(Shandong Yunhai Guochuang Cloud Computing Equipment Industry Innovation Co.,Ltd.,Jinan 250013,China;Shandong Inspur Artificial Intelligence Research Institute Co.,Ltd.,Jinan 250013,China;School of Computer Science&Technology,Shandong Jianzhu University,Jinan 250101,China)

机构地区：[1]山东云海国创云计算装备产业创新中心有限公司,济南250013 [2]山东浪潮人工智能研究院有限公司,济南250013 [3]山东建筑大学计算机科学与技术学院,济南250101

出　　处：《计算机应用研究》2025年第3期663-676,共14页Application Research of Computers

基　　金：山东省自然科学基金青年基金资助项目(ZR2023QF050,ZR2023QF056);国家自然科学基金资助项目(62176141);山东省自然科学基金杰出青年基金资助项目(ZR2021JQ26)。

摘　　要：视频异常事件检测逐渐成为计算机视觉领域的研究热点之一,具有重要研究意义和应用价值。近年来,以卷积神经网络为核心的深度学习技术在多项机器视觉任务中展现优异性能,极大地启发了其在视频异常事件检测领域的应用。为此,对近年来基于深度学习的视频异常事件检测相关研究进行全面梳理与系统归纳。首先,根据视频异常检测实现流程的三个核心要素,即检测模式、样本设置及学习/推理机制,提出一种由浅入深的多级分类方案,面向前沿深度学习方法开展逐类概述并提炼代表性算法数学模型,同时聚焦现有方法的局限性进行阐述;其次,介绍本领域主流的基准测试数据集,汇总并对比当前先进方法在不同数据集上的检测性能;最后,围绕复杂光照/天气条件、多模态图像显著融合、可语义解释及自适应场景感知四个方面对未来重点研究方向进行讨论和展望,期望为该领域的后续研究提供借鉴与参考。VAD is one of the hottest research topics in the field of computer vision,which is significant for research and valuable for application.In recent years,inspired by the outstanding performances of deep learning technologies represented by the convolution neural networks on various tasks of machine vision,a large number of deep learning-based VAD researches have rapidly emerged.To this end,this paper comprehensively sorted out and systematically summarized the deep learning-based VAD researches.Firstly,it proposed a multi-level classification scheme based on the three core elements of anomaly detection process including detection strategy,sample setting and learning/inferring mechanism,which was utilized to summarize the frontier deep learning-based VAD methods by class,refined the mathematical models of representative algorithms and elaborated the limitations of existing works simultaneously.Secondly,it introduced the benchmark datasets of video anomaly detection and compared the performances of advanced methods on diverse datasets.Finally,it discussed the future research directions in four aspects as follows:complex lighting/weather conditions,fusion of multi-modal images,semantic interpretability and adaptive scene perception,which was expected to provide references for future research works in this field.

关 键 词：智能监控 视频异常检测 深度学习 卷积神经网络 生成对抗网络 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]