1 School of Information Network Security, People’s Public Security University of China, Beijing 100038, China
2 Graduate School, People’s Public Security University of China, Beijing 100038, China
Show less
文章历史+
Received
Published
Revised
2025-11-10
2026-10-28
2026-04-21
Online Date
2026-11-10
PDF (7489K)
摘要
为解决视频异常检测现有方法在细粒度视频异常事件预测中表现不足的问题,提出了一种基于对比语言-图像预训练模型(contrastive language-image pre-training,CLIP)的多模态特征融合的细粒度视频异常检测方法(CLIP-based multimodal feature fusion for fine-grained video anomaly detection,ClipFusionVAD)。首先,利用对比语言-图像预训练模型分别提取视频帧与文本特征,设计特征聚合器模块和特征投影器模块实现不同维度上的特征融合,提升多模态特征表达能力,实现准确的细粒度异常事件分类。然后,引入结合通道注意力机制和空间注意力机制的全局注意力模块(global attention module,GAM),增强视觉特征的判别能力。实验结果表明,ClipFusionVAD在细粒度异常检测任务中表现优异,在UCF-Crime和XD-Violence数据集上mAP分别达到9.4%和28.07%,进一步验证了该方法的有效性。
Abstract
To address the limitations of existing video anomaly detection methods in fine-grained anomaly event prediction, a CLIP-based multimodal feature fusion method for fine-grained weakly supervised video anomaly detection, named ClipFusionVAD, was proposed. First, features from video frames and text were extracted separately using the pre-trained CLIP model. A feature aggregator and a feature projector were designed to achieve multi-dimensional feature fusion, enhancing the representation capability of multimodal features and enabling accurate fine-grained anomaly event classification. Furthermore, GAM (global attention module), which combines channel and spatial attention mechanisms, was introduced to improve the discriminative power of visual features. Experimental results demonstrate that ClipFusionVAD achieves outstanding performance on fine-grained anomaly detection tasks, with mAP attaining 9.4% on the UCF-Crime dataset and 28.07% on the XD-Violence dataset, which further verifies the effectiveness of the proposed method.
视频异常检测通过自动识别异常行为,提升安防系统的响应效率和准确性[1-2],在智能监控和公共安全中至关重要,广泛应用于公共场所监控、智能交通管理及高安全区域的入侵防护,帮助安保人员快速反应,预防潜在威胁,是现代安防体系中不可或缺的核心工具,为维护社会公共安全提供了强有力的技术支持[3]。视频异常检测根据监督方式可分为弱监督[4]、半监督[5]、全监督[6]和无监督[7]4类。全监督方法依赖精确标注数据,检测准确率高,但在大规模应用中成本较大且难以扩展;无监督方法无需标注数据,灵活性较强,但性能较弱。半监督方法通过利用大量标注的正常数据和少量异常数据减少标注成本并提升检测效果,但对标注数据的依赖仍限制了此类方法的应用场景。弱监督视频异常检测(weakly supervised video anomaly detection,WSVAD)只需视频级标签,能够更好地平衡标注成本与检测性能,显著降低标注成本的同时保持良好的检测效果。
细粒度视频异常检测是在实现粗粒度异常事件检测的基础上进一步识别异常事件的具体类型。通过关注视频中的局部特征,区分不同异常事件的视觉特征差异与语义边界,进而判断视频中发生的异常类别,实现异常事件的细分类型识别。Li等[16]引入3D卷积神经网络(coordinate 3D,C3D)通过时空卷积捕捉空间和时间特征,但在长时序视频的处理上表现有限。Feng等[17]提出自引导时间辨别Transformer,虽然增强了时序依赖的建模能力,但在背景复杂或异常行为变化迅速的情况下,其定位精度仍需改进。An等[18]提出自适应特征细化(aggregation with fragment retransmission,AFR)方法,虽然能够自适应地增强关键区域的特征表示,但是对极端场景与小尺度异常的鲁棒性不足。Zhou等[19]提出基于运动和外观引导的补丁扩散模型进行视频异常检测,有效提升了对小尺度异常和复杂行为的微观区分能力,但是其推理速度较慢,实时性不足。
粗粒度视频异常检测:在XD-Violence数据集中,使用帧级平均精度(average precision,AP)作为评估指标。AP值通过计算不同阈值下的精度-召回曲线下的面积来确定。在UCF-Crime数据集中,使用帧级曲线下面积(area under the curve,AUC)作为评估指标。AUC用于衡量模型在不同阈值下区分正常帧和异常帧的能力。
细粒度弱监督视频异常检测:使用的指标包括平均精度(mean average precision,mAP)和在不同交并比(intersection over union,IoU)阈值下的mAP平均值(average,AVG)。mAP表示每个细分类别异常事件的平均精度值,反映了模型在检测不同类别异常时的整体性能。交并比(IoU)用于衡量预测的异常时间段与真实异常时间段之间的重叠程度。
2.1.3 参数设置
ClipFusionVAD模型基于预训练的CLIP模型(ViT-B/16)构建,并使用冻结的图像和文本编码器作为网络架构的基础。模型的前馈网络采用标准的Transformer层,并将激活函数从ReLU替换为GELU。在超参数设置方面,位置距离部分的σ值设为1,多分类预测中的温度参数τ设为0.07,可学习提示的上下文长度为20。XD-Violence数据集的特征提取视觉序列长度为64帧,而UCF-Crime数据集为8帧。总损失函数中的权重因子γ 在两个数据集中分别设置为1×10-4和1×10-1。模型使用PyTorch在NVIDIA RTX 2080 Ti GPU上进行训练,优化器采用AdamW,批量大小设为64。学习率在XD-Violence数据集中设为2×10-5,UCF-Crime数据集中设为1×10-5,训练轮次均为10。
ZhangLin, ChenZhaobo, MaXiaoxuan, et al. Review of unsupervised and weakly super vised video anomaly detection methods[J]. Computer Engineering and Applications, 2024, 24(19): 7941-7955.
[3]
ZhuS, ChenC, SultaniW. Video anomaly detection for smart surveillance[M]// Computer Vision:A Reference Guide. Cham: Springer International Publishing, 2021: 1315-1322.
[4]
NayakR, PatiU C, DasS K. A comprehensive review on deep learning-based methods for video anomaly detection[J]. Imageand Vision Computing, 2021, 106: 104078.
[5]
LiuY, LiuJ, ZhaoM, et al. Collaborative normality learning framework for weakly supervised video anomaly detection[J]. IEEE Transactions on Circuits and Systems II: Express Briefs, 2022, 69(5): 2508-2512.
[6]
HuangC, LiuC, WenJ, et al. Weakly supervised video anomaly detection via self-guided temporal discriminative transformer[J]. IEEE Transactions on Cybernetics, 2022, 54(5): 3197-3210.
[7]
HasanM, ChoiJ, NeumannJ, et al. Learning temporal regularity in video sequences[C]// Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Las Vegas: IEEE, 2016: 733-742.
[8]
ZaheerM Z, MahmoodA, KhanM H, et al. Generative cooperative learning for unsupervised video anomaly detection[C]// Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. New Orleans: IEEE, 2022: 14744-14754.
[9]
UllahW, UllahA, HaqI U, et al. CNN features with bi-directio-nal LSTM for real-time anomaly detection in surveillance networks[J]. Multimedia Tools and Applications, 2021, 80: 16979-16995.
[10]
ZhongJ X, LiN, KongW, et al. Graph convolutional label noise cleaner: train a plug-and-play action classifier for anomaly detection[C]// Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. Long Beach: IEEE, 2019: 1237-1246.
ChenCheng, HuYan. Video anomaly detection combining memory-augmented[J]. Computer Engineering and Applications, 2022, 58(15): 253-259.
[13]
FanY, YuY, LuW, et al. Weakly-supervised video anomaly detection with snippet anomalous attention[J]. IEEE Transactions on Circuits and Systems for Video Technology, 2024(7): 5480-5492.
XuChenyue, WangRong, GuoFang, et al. Abnormal behavior recognition method of improved spatio-temporal graph normalizing flow[J]. Science Technology and Engineering, 2025, 25(18): 7693-7699.
[20]
Al-lahhamA, TastanN, ZaheerM Z, et al. Acoarse-to-fine pseudo-labeling (C2FPL) framework for unsupervised video anomaly detection[C]// Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). Waikoloa: IEEE, 2024: 6779-6788.
[21]
LiJ, LiuX, ZhangW, et al. Spatio-temporalattention networks for action recognition and detection[J]. IEEE Transactions on Multimedia, 2020, 22(11): 2990-3001.
[22]
FengX, SongD, ChenY, et al. Convolutional transformer based dual discriminator generative adversarial networks for video anomaly detection[C]// Proceedings of the 29th ACM International Conference on Multimedia. Chengdu: ACM, 2021: 5546-5554.
[23]
AnD, LiuW, LuJ, et al. Fine-grained video anomaly detection via adaptive feature refinement and semantic enrichment[J]. Visual Computer, 2026, 42(2): 1-18.
[24]
ZhouH, CaiJ, YeY, et al. Video anomaly detection with motion and appearance guided patch diffusion model[C]// Proceedings of the AAAI Conference on Artificial Intelligence (AAAI). Philadelphia: AAAI, 2025: 10761-10769.
[25]
RadfordA, KimJ W, HallacyC, et al. Learning transferable vi-sual models from natural language supervision[C]// International Conference on Machine Learning. Online: PMLR, 2021: 8748-8763.
[26]
JuC, HanT, ZhengK, et al. Prompting visual-language models for efficient video understanding[C]// European Conference on Computer Vision. Cham: Springer Nature Switzerland, 2022: 105-124.
[27]
JooH K, VoK, YamazakiK, et al. CLIP-TSA: CLIP-Assisted temporal self-attention for weakly-supervised video anomaly detection[C]// IEEE International Conference on Image Processing (ICIP). Kuala Lumpur: IEEE, 2023: 3230-3234.
[28]
WuP, ZhouX, PangG, et al. Vadclip: adapting vision-language models for weakly supervised video anomaly detection[J]. Proceedings of the AAAI Conference on Artificial Intelligence, 2024, 38(6): 6074-6082.
[29]
LüH, YueZ, SunQ, et al. Unbiased multiple instance learning for weakly supervised video anomaly detection[C]// Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. Vancouver: IEEE, 2023: 8022-8031.
[30]
LiuL, LiJ, LiG, et al. VadCLIP++: dynamic vision-language model for weakly supervised video anomaly detection[J]. Digital Signal Processing, 2026, 168: 105560.
[31]
KimJ, YoonS, ChoiT, et al. Unsupervised video anomaly detection based on similarity with predefined text descriptions[J]. Sensors, 2023, 23(14): 6256.
[32]
SultaniW, ChenC, ShahM. Real-world anomaly detection in surveillance videos[C]// Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Salt Lake City: IEEE, 2018: 6479-6488.
[33]
WuP, LiuJ, ShiY, et al. Not only look, but also listen: learning multimodal violence detection under weak supervision[C]// Computer Vision-ECCV 2020: 16th European Conference. Glasgow: Springer International Publishing, 2020: 322-339.
[34]
SchölkopfB, WilliamsonR C, SmolaA, et al. Support vector method for novelty detection[J]. Advances in Neural Information Processing Systems, 1999: 582-588.
[35]
ZhouH, YuJ, YangW. Dual memory units with uncertainty re-gulation for weakly supervised video anomaly detection[C]// Proceedings of the AAAI Conference on Artificial Intelligence. Was-hington, D. C. : AAAI, 2023: 3769-3777.
[36]
TianY, PangG, ChenY, et al. Weakly-supervised video anomaly detection with robust temporal feature magnitude learning[C]// Proceedings of the IEEE/CVF International Conference on Computer Vision. Montreal: IEEE, 2021: 4975-4986.