We propose NL-VRS, a hierarchical framework for retrieving abnormal events in large-scale indoor CCTV archives using natural language queries. The system employs a hierarchical two-stage framework: fast candidate retrieval using domain-adapted CLIP (C...
We propose NL-VRS, a hierarchical framework for retrieving abnormal events in large-scale indoor CCTV archives using natural language queries. The system employs a hierarchical two-stage framework: fast candidate retrieval using domain-adapted CLIP (Contrastive Language-Image Pretraining) and FAISS (Facebook AI Similarity Search), followed by structured metadata generation via a video VLM (Vision-Language Model), LLaVA-NeXT-Video-7B. We bridge the domain gap using LoRA (Low-Rank Adaptation)-based fine-tuning on the CLIP visual encoder and incorporate adaptive temporal sampling to capture behavior-rich intervals while minimizing frame redundancy. Experiments on the AI Hub indoor abnormal behavior dataset show that the proposed method improves Recall@1 from 33.5% to 41.5%, Recall@5 from 42.0% to 76.5%, and conditional mIoU (mean Intersection over Union, calculated only when the top-1 result satisfies the category match condition) from 0.348 to 0.409 over the baseline CLIP-based retrieval system. The first-stage retrieval takes 57 ms on average, while the full pipeline requires about 2.1 seconds, demonstrating the practical feasibility of the proposed framework for real-time surveillance support.