Using Vision-Language Models to Detect Anomalies in CCTV Surveillance Videos

View at publisher Summarise with VIRAT
Authors: Jiho Choi, Sung-Woo Byun, Jong Seung Park, Seung Pyo Hong
Cited by: 0
References: 10

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Real-time detection of anomalous behaviors in CCTV footage remains challenging due to the rarity of incidents, varied environments, and class imbalance. While large Vision-Language Models like CLIP have demonstrated robust image understanding, they lack temporal modeling needed for video anomaly recognition. In this work, we propose a temporal-aware CLIP-based framework to detect abnormal events in surveillance video. The model integrates a lightweight temporal transformer and a multimodal fusion mechanism that combines visual features with class-specific text embeddings, enabling context-aware video understanding. To address the scarcity of rare events, we augment the training data with synthetic video clips generated using OpenAI’s Sora text-to-video model. Experiments on a public CCTV anomaly dataset show that our enhanced CLIP model outperforms a baseline image-level CLIP classifier, with substantial improvements in precision, recall, and F1-score. In particular, synthetic augmentation boosted the detection of infrequent behaviors like "fight" and "fall down." These results demonstrate the practical benefit of combining pre-trained vision-language representations with generative data augmentation for anomaly detection. We also discuss the contributions of each component and remaining challenges.

Fields of Study & Subjects

Anomaly Detection Techniques and Applications Multimodal Machine Learning Applications Human Pose and Action Recognition

Publication details

Journal / source
International Conference on Consumer Electronics
Pages
1–6
Publication date
3 February 2026
Language
EN