Testing Video Anomaly Detection through State Space Models on Challenging Real-World Surveillance Datasets
Esther González Amorós, Alberto Chueco, Míriam Máximo, Carlos Viegas, Adrián Peidró, Luis Payá
23rd International Conference on Informatics in Control, Automation and Robotics - ICINCO 2026  (Angers, Francia, 26-28 octubre 2026)
Ed. SciTePress  ISBN:978-989-758-870-9  ISSN:2184-2809  DOI:10.5220/0015459500004979  BIBTEX:@conference{icinco26, author={Esther González{-}Amorós and Alberto Chueco and Míriam Máximo and Carlos Viegas and Adrián Peidró and Luis Paya}, title={Testing Video Anomaly Detection through State Space Models on Challenging Real-World Surveillance Datasets}, booktitle={Proceedings of the 23rd International Conference on Informatics in Control, Automation and Robotics - Volume 1: ICINCO}, year={2026}, pages={639-648}, publisher={SciTePress}, organization={INSTICC}, doi={10.5220/0015459500004979}, isbn={978-989-758-870-9}, issn={2184-2809}, }  - Volume 1 - pages 639-648

Resumen:

This work presents the first comprehensive evaluation of the VADMamba framework on the UCF-Crime dataset, a challenging benchmark for video anomaly detection (VAD) in real-world surveillance environments. While VADMamba has demonstrated strong performance on controlled-environment benchmarks, its applicability to semantically complex surveillance scenarios remains unexplored. Through systematic experimentation, we demonstrate that a lightweight, prediction-only configuration utilizing a depth profile of [3,3,3,3] achieves an AUC (Area Under Curve) of 72.23% and 1.8 times fewer parameters and 2.6 times fewer FLOPs than the full original dual-task pipeline, by eliminating the optical flow reconstruction bottleneck. This performance surpasses the previous state-of-the-art one-class classification methods on UCF-Crime, despite training on only 260 videos of the available 800 normal videos. Cross-dataset transfer experiments reveal substantial domain gaps when models trained on controlled benchmarks are evaluated on UCF-Crime, highlighting the limited representativeness of widely-used datasets for real-world deployment.