Publication Type
Journal Article
Version
acceptedVersion
Publication Date
3-2027
Abstract
Video Anomaly Understanding (VAU) aims to identify, localize, and interpret anomalous events in videos. While foundation models have significantly advanced VAU in recent years, a comprehensive review dedicated to this emerging field remains underexplored. To bridge this gap, this paper presents a systematic survey of foundation model-driven VAU, providing a structured analysis of representative methods from a unified anomaly perception-cognition perspective. Specifically, the anomaly perception focuses on exploiting the representation, alignment, and generalization capabilities of foundation models to discover and localize potential anomalous cues from videos. The anomaly cognition further leverages reasoning and generative abilities to support semantic interpretation, enabling a transition from anomaly detection to genuine anomaly understanding. Under this unified framework, we clearly characterize the differences among existing methods in terms of their capability levels and reveal the central role of foundation models in bridging low-level anomaly perception and high-level anomaly cognition. Building upon this analysis, we also review the evolution of VAU benchmarks and evaluation metrics, highlighting the shift from coarse-grained classification scoring toward fine-grained semantic understanding. Finally, we further discuss the key challenges and future directions in terms of fast-slow collaboration, long-form understanding, synthetic video assistance, multi-view exploration, and unified benchmarks and evaluations for VAU.
Keywords
Video anomaly detection, Video anomaly understanding, Foundation models, Vision-language models
Discipline
Artificial Intelligence and Robotics | Graphics and Human Computer Interfaces
Research Areas
Intelligent Systems and Optimization
Areas of Excellence
Digital transformation
Publication
Pattern Recognition
Volume
183
First Page
1
Last Page
12
ISSN
0031-3203
Identifier
10.1016/j.patcog.2026.114890
Publisher
Elsevier
Citation
SU, Wanshun; TANG, Ke; LIU, Dingyi; WU, Peng; YAN, Qingsen; and PANG, Guansong.
A survey on foundation model-driven video anomaly understanding. (2027). Pattern Recognition. 183, 1-12.
Available at: https://ink.library.smu.edu.sg/sis_research/11411
Creative Commons License
This work is licensed under a Creative Commons Attribution-NonCommercial-No Derivative Works 4.0 International License.
Additional URL
https://doi.org/10.1016/j.patcog.2026.114890
Included in
Artificial Intelligence and Robotics Commons, Graphics and Human Computer Interfaces Commons