Confidence-Aware Cascaded Learning for Real-Time Weakly Supervised Anomaly Detection in Surveillance Video
No Thumbnail Available
Date
2026-04-01
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Institute of Electrical and Electronics Engineers Inc.
Abstract
Weakly supervised video anomaly detection (WS-VAD) has advanced significantly, yet most approaches remain tailored to long, untrimmed streams and fail to meet the latency and reliability demands of real-world surveillance. In asset-protection scenarios, where motion-triggered clips must be processed in real-time, existing methods often compromise either accuracy or efficiency. We present a Confidence-Aware Sequential Routing Ensemble (CASRE), a cascaded framework designed for Clip-Level Weak Supervision (CLWS). CASRE integrates a temperature-scaled ResNet3D (TDR3D) primary detector with a novel Temporal-Scale Delta Loss to enforce confidence separation. Only low-confidence clips are routed to a transformer-based MViTv2 fallback, and final predictions are obtained through calibrated probabilistic fusion. This selective routing ensures that high-capacity models are invoked for fewer than 6% of clips, while the majority benefit from lightweight inference. Extensive evaluation on UCF–Crime, UCSD–Ped1/Ped2, and the newly introduced SSAD dataset, a real-world benchmark curated for asset-protection surveillance, demonstrates that CASRE achieves up to 98.64% accuracy and 99.91% AUC, while maintaining sub-0.03 s inference for most samples. By combining confidence calibration, selective routing, and heterogeneous backbone fusion, CASRE establishes a deployable WS-VAD paradigm that balances accuracy, robustness, and computational efficiency, making it well-suited for large-scale surveillance deployments under strict operational constraints.