Metric-Wise Comparative Analysis of Hybrid CNN–SRU/LSTM and Lightweight CNN–MIL Frameworks for Deployment-Oriented Video Anomaly Detection

Abstract

Video anomaly detection is a critical component of intelligent surveillance systems, where detection accuracy, temporal stability, computational efficiency, and real-world deployment feasibility must be jointly considered. Existing studies frequently rely on ROC–AUC as the primary evaluation metric, providing limited insight into practical system performance. This study presents a structured metric-wise comparative analysis of hybrid CNN–SRU/LSTM architectures and lightweight CNN-based multiple instance learning (MIL) frameworks, based on systematically collected benchmark results from datasets such as UCF-Crime and ShanghaiTech. The analysis follows a literature-driven methodology and evaluates models across multiple dimensions, including AUC, false alarm rate (FAR), temporal stability, inference speed (FPS), computational footprint, and calibration reliability using expected calibration error (ECE). Deployment-oriented factors such as latency–performance trade-offs, cross-domain robustness, and scalability under limited labeled data are also examined. Results indicate that hybrid CNN–SRU/LSTM frameworks achieve approximately 85.9% AUC across benchmark datasets with strong temporal consistency, while CNN–MIL approaches maintain competitive accuracy (≈82–84.7%) with significantly higher efficiency (up to 72 FPS) and improved calibration (ECE reduced from ~0.17 to ~0.10). Transformer-based and vision–language models achieve slightly higher accuracy (>86% AUC) but operate at substantially lower frame rates (<12 FPS) and higher memory requirements (>800 MB). These findings highlight that marginal accuracy gains often incur substantial computational cost, emphasizing multi-metric evaluation and hardware-aware model selection for practical video anomaly detection systems.