Confidence-Calibrated Temporal Attention Framework for Interpretable Video Anomaly Detection

Abstract Video anomaly detection in surveillance environments is challenged by ambiguous event boundaries, complex scenes, and unreliable predictions. Existing methods mainly focus on anomaly scoring without modeling prediction confidence or interpretability. This paper proposes a confidence-guided spatiotemporal framework for interpretable anomaly detection. The method introduces a confidence calibrated temporal attention mechanism to jointly model anomaly likelihood and prediction reliability, enabling distinction between reliable and uncertain detections. It further employs uncertainty-aware attention to adaptively aggregatetemporal features for robust anomaly localization. A confidence estimation module captures uncertainty, while temporal consistency stabilizes predictions. An interpretability module highlights salient spatiotemporal regions contributingto anomaly decisions. Experiments on UCSD, Avenue, ShanghaiTech, and UCFCrime datasets achieve 95.1%, 89.3%, 87.3%, and 84.7% AUC, respectively, with an average improvement of 2.8% over baselines and 58 FPS inference speed. These results demonstrate that confidence-aware and interpretable modeling improves the reliability and applicability of video anomaly detection systems.