Пошук небезпечних подій у відеоархівах спостереження за текстовим описом на основі моделей глибокого навчання

Abstract

This paper explores the application of vision-language alignment models for prompt-based retrieval of dangerous events in surveillance video archives. Traditional anomaly detection requires extensive labeled data and retraining for each scenario. The proposed approach leverages contrastive pretraining (CLIP, InternVideo2) and LLM-driven anomaly scoring (LAVAD, Holmes-VAD) to enable natural-language retrieval of hazardous events without fine-tuning. We analyze AnomalyCLIP's zero-shot anomaly recognition and Holmes-VAD's explainable detection on UCF-Crime, XD-Violence, and UCA benchmarks. A cascaded architecture combining lightweight CLIP-scoring with LLM-driven semantic verification is proposed for scalable, interpretable surveillance.