Resource-Aware Confidence-Oriented Late Fusion for Unsupervised Audio–Visual Traffic Anomaly Detection

Unsupervised audio–visual anomaly detection has emerged as a promising approach for intelligent traffic monitoring, particularly in scenarios where anomalous events are rare, diverse and difficult to annotate exhaustively. Existing methods prioritize peak detection accuracy under fixed thresholds, while overlooking deployment-oriented concerns such as score stability, threshold sensitivity and robustness under realistic operating conditions. This paper presents RACoF, a Resource-Aware Confidence-Oriented Fusion framework for unsupervised audio–visual traffic anomaly detection. Instead of tightly coupling multimodal feature learning, RACoF decouples modality-specific anomaly scoring from fusion and decision-making. Audio and visual anomaly detectors are trained independently using normal-only data. Their scores are subsequently normalized, fused and calibrated through a validation driven percentile thresholding strategy. This modular score-level or resource-aware approach mitigates scale mismatch across modalities and reduces sensitivity to absolute score magnitudes with negligible fusion overhead. Extensive experiments on real-world traffic datasets demonstrate that, although the proposed fusion strategy does not consistently outperform the strongest unimodal video detector in terms of peak AUC (Area Under Curve), it yields significantly more stable decision regions, lower threshold sensitivity, and improved interpretability across varying operating regimes. Further analysis of threshold behavior, alarm distributions, and regime-dependent fusion weights highlights RACoF’s suitability for deployment-oriented traffic monitoring systems. Importantly, RACoF is model-agnostic and supports lightweight configurations by substituting heavy backbones with mobile-friendly audio–visual models, making it compatible with resource-constrained edge platforms. These results suggest that emphasizing decision stability and calibration, rather than peak accuracy alone, provides a practical pathway toward robust and edge deployable multimodal traffic anomaly detection.