SF-VAD: Scene-Aware Video Anomaly Detection via Projected Semantic Context

Video anomaly detection in multi-scenario surveillance is challenging because normal activity patterns vary dramatically across environments, making it harder to distinguish anomalies from legitimate activity. Existing weakly-supervised methods treat all scenarios with a single global model, ignoring this contextual prior. We propose SF-VAD, a scene-aware framework that enriches motion-centric I3D features with semantic scene context derived from a learned I3D-to-CLIP projection. Our approach (i) trains a projection network on a small set of paired I3D-CLIP features to produce pseudo-CLIP embeddings for every video, and (ii) injects the projected scene context additively into a temporal transformer encoder. Under the standard weakly-supervised protocol, SF-VAD achieves 88.27% frame-level AUC on the MSAD benchmark, closely approaching the multimodal pi-VAD (88.47%) while surpassing all existing I3D-based methods and requiring only two feature types compared to six modalities.