A Deployment-Oriented MCDA Framework for Selecting Industrial Video Anomaly Detection Architectures

Video Anomaly Detection (VAD) has emerged as a promising technology for automated surveillance and industrial monitoring. However, selecting an appropriate VAD architecture for real-world deployment remains challenging because benchmark-oriented evaluations often overlook practical considerations such as computational efficiency, latency, maintainability, supervision requirements, and long-term operational stability. To address this gap, this study proposes an expert-driven Multi-Criteria Decision Analysis (MCDA) framework for the deployment-oriented selection of VAD architectures. Eight representative architecture families were evaluated against six industrially relevant criteria, with scenario-specific priorities derived using the Best–Worst Method (BWM) from five domain experts across four representative industrial deployment scenarios. Edge-oriented architectures ranked first in three scenarios, achieving MCDA scores of 4.26, 3.90, and 4.07, whereas lightweight CNN-based architectures achieved the highest score (4.12) in the resource-constrained scenario. Inter-expert agreement ranged from Kendall’s W = 0.54 to 0.85, and Monte Carlo analysis confirmed the robustness of rankings, with top-rank probabilities of 69–74% for edge-oriented architectures and 100% for lightweight CNNs in the resource-constrained scenario. These findings demonstrate that architectural suitability depends on the deployment context rather than on a universally superior modeling paradigm, and that industrial VAD should be approached as a deployment-oriented systems engineering problem. The proposed framework provides a transparent and robust basis for aligning VAD architecture selection with operational requirements.