【作者】 李博；

【导师】 葛宏伟；

【作者基本信息】 大连理工大学 ， 计算机科学与技术， 2025， 硕士

【摘要】 视频异常检测任务从视频中自动识别异常对象或异常行为,由于异常事件的稀缺性和异常边界的模糊性,这一任务极具挑战。传统的视频异常检测方法通常依赖简单的手工特征,缺乏高级语义信息的表达,而基于深度学习的方法则展现了强大的潜力。现有的深度学习方法大多单独使用重建或预测代理任务约束模型,或融合其他自监督代理任务进行多任务学习,以充分发挥不同代理任务各自的优势。然而,多任务学习虽然能够提供更全面的指导,但也可能因任务间的不适配和缺乏关联导致模型陷入局部最优。此外,多任务学习在真实场景下客观存在训练部署流程繁琐和计算资源消耗大的固有问题。为此,本文针对上述两方面的局限性,基于单类分类学习和无监督学习,围绕视频异常检测中代理任务的优化和创新展开研究,提出了两种新颖的视频异常检测方法,主要研究内容和贡献如下:(1)针对多任务学习中子代理任务之间不适配和缺乏关联的问题,本文提出了一种基于联合多任务分层对比学习的视频异常检测算法(Joint Multi-Task Hierarchical Contrastive Learning,JMHCL)。首先,设计了一种外观-运动特征自适应交互融合模块,通过精细的通道级注意力机制和自适应融合策略,实现RGB图像特征与运动特征的深度协同融合,减少特征冗余并为后续多任务学习提供更具表征性和判别性的特征表示。其次,提出了一种多代理任务学习模块,通过缺失帧位置预测、帧顺序预测和中间帧预测三个互补任务,深入挖掘视频序列中潜在的语义关联和细粒度时空变化,显著增强模型对视频时空信息的理解能力。最后,引入联合代理任务的双层对比学习机制,通过样本到样本和样本到原型的对比学习,不仅能够放大不同样本间的差异性,还能通过原型学习增强正常样本表示,有效平衡多任务学习中的特征的关联性和一致性。实验结果表明,JMHCL在三个公共数据集上表现出优异的性能,整体结果显著优于现有的单任务和多任务方法。(2)针对多任务学习存在知识负迁移、模型设计复杂和训练调参困难的问题,本文提出了一种基于无监督多尺度恢复扩散模型的视频异常检测方法,(Multiscale Re-covery Diffusion Model,MRDM)。首先,通过数学推导和实验验证设计了一种基于扩散模型的恢复代理任务,有效扩大正常与异常样本分布之间的差异。具体而言,基于扩散模型的生成特性和视频异常检测的基本假设,本文证明了正常事件的生成过程比异常事件更快,从而在关键去噪步骤中实现更高效的异常检测。其次,提出了一种多尺度时空差分模块,通过轻量化的卷积和逐元素减法操作,实现对前景物体多尺度外观和运动信息的提取,增强模型对事件前景关注度的同时减少对预训练模型的依赖。最后,引入特征分布推拉模块,提出了一种全新的“推拉”损失函数,通过在训练过程中迭代学习和更新伪标签,使模型对正常样本去噪噪声的预测能力进一步优于异常样本,从而扩大正常与异常样本的分布差异。实验结果表明,MRDM在五个公共数据集上取得了显著的性能提升,尤其在真实场景下异常检测任务的综合表现具有突出优势。更多还原

【Abstract】 Video anomaly detection aims to automatically identify abnormal objects or behaviors in videos,a highly challenging task due to the scarcity of anomalous events and the ambigu-ity of anomaly boundaries.Traditional methods typically rely on simple handcrafted features lacking high-level semantic expression,while deep learning-based approaches demonstrate greater potential.Most existing deep learning methods either employ reconstruction or pre-diction proxy tasks individually,or integrate other self-supervised proxy tasks for multi-task learning to leverage their respective advantages.However,although multi-task learning can provide more comprehensive guidance,it may lead to suboptimal solutions due to task incom-patibility and insufficient inter-task correlations.Moreover,multi-task learning inherently suffers from complex training,deployment processes and substantial computational resource consumption in real-world scenarios.To address these limitations,this paper focuses on optimizing and innovating proxy tasks for video anomaly detection based on one-class clas-sification and unsupervised learning,proposing two novel video anomaly detection methods.The main research contributions are as follows:(1)To resolve task incompatibility and insufficient correlations in multi-task learning,we propose a Joint Multi-Task Hierarchical Contrastive Learning(JMHCL)algorithm.First,an appearance-motion feature adaptive interaction fusion module is designed,achieving deep collaborative fusion of RGB image features and motion features through channel-level atten-tion mechanisms and adaptive fusion strategies,reducing feature redundancy while providing more representative and discriminative features.Second,a multi-proxy task learning module is developed,employing three complementary tasks(missing frame localization prediction,frame order prediction,and intermediate frame prediction)to thoroughly explore latent se-mantic correlations and fine-grained spatiotemporal variations in video sequences,signifi-cantly enhancing the model’s understanding of spatiotemporal information.Finally,a dual-level contrastive learning mechanism combining sample-to-sample and sample-to-prototype contrast is introduced,which not only amplifies inter-sample differences but also enhances normal sample representations through prototype learning,effectively balancing feature rele-vance and consistency in multi-task learning.Experimental results demonstrate that JMHCL achieves superior performance on three public datasets,significantly outperforming existing single-task and multi-task methods.(2)To address knowledge negative transfer,complex model design,and training difficul-ties in multi-task learning,we propose an unsupervised Multiscale Recovery Diffusion Model(MRDM).First,a diffusion model-based recovery proxy task is mathematically derived and experimentally validated to effectively amplify distribution differences between normal and abnormal samples.Specifically,based on the generative characteristics of diffusion models and fundamental assumptions of video anomaly detection,we demonstrate that normal events complete their generation process faster than anomalies,enabling more efficient anomaly detection at critical denoising steps.Second,a lightweight multiscale spatiotemporal sub-traction module is developed using convolutional operations and element-wise subtraction to extract multi-scale appearance and motion information of foreground objects,enhancing event foreground attention while reducing dependency on pre-trained models.Finally,a fea-ture distribution push-pull module with a novel push-pull loss function is introduced,which iteratively learns and updates pseudo-labels during training to further improve the model’s noise prediction capability for normal samples compared to abnormal ones,thereby enlarging their distribution discrepancy.Experimental results show that MRDM achieves remarkable performance improvements on five public datasets,demonstrating particularly outstanding advantages in real-world anomaly detection scenarios.更多还原