StreamVAD: A streaming framework with progressive context integration for multi-temporal scale video anomaly detection

Video anomaly detection (VAD) plays a crucial role in intelligent surveillance systems by identifying abnormal events in video streams. However, most existing methods either rely on isolated feature extraction—failing to model inter-action contextual relationships critical for complex anomaly recognition—or demand full-video processing via graph/hierarchical architectures, which incur high latency, computational burden, and parameter/memory inefficiency with depth. Lightweight designs mitigate costs but sacrifice temporal sensitivity through shallow networks and short-clip inputs, limiting detection of subtle or multi-scale anomalies in streaming scenarios. To address these challenges, we propose StreamVAD, a lightweight streaming anomaly detection framework that achieves low-latency, long-term temporal modeling with minimal computational overhead. A Key Clip Generator (KCG) filters redundant inputs in a streaming manner, allowing the model to focus on informative content while reducing computational cost. A progressive context integration (PCI) module incrementally expands the temporal receptive field by integrating historical context without full-sequence buffering, enabling efficient detection of complex long-term anomalies. Additionally, a multi-scale temporal selection (MTS) strategy dynamically adapts temporal resolution to capture both short- and long-term abnormalities. Extensive experiments on UCF-Crime, XD-Violence, and a supplemental long-term anomaly dataset demonstrate that StreamVAD achieves effective video anomaly detection with fewer parameters and lower latency. The code and dataset are available at https://github.com/Han-lijun/StreamVAD.