VADOR: Real World Video Anomaly Detection with Object Relations and Action
Halil İbrahim Öztürk (Togg),* Ahmet Burak Can (Hacettepe University)
The 34th British Machine Vision Conference
Abstract
Anomaly detection in real world videos requires complex scene understanding. Previous works utilize action recognition models as feature extractor, but some anomalies (e.g. robbery) can not be easily understood from basic action information. Our VADOR model leverages action and relationships of objects in the scene to detect anomaly using transformer encoders. Cross-attention between object relation encoder and action encoder helps to fusion of information. Our Anchor based Temporal Action Localization network (TALNet) segments anomalies temporarily by using clip features generated from the encoders. We train VADOR with strong regularization and data augmentation methods. VADOR achieves \%83.62 AUC score while achieving \%63.09 F1@25 score at temporal segmentation on UCF Crime dataset. Code is publicly available at https://github.com/hibrahimozturk/vador.Video
Citation
@inproceedings{Öztürk_2023_BMVC,
author = {Halil İbrahim Öztürk and Ahmet Burak Can},
title = {VADOR: Real World Video Anomaly Detection with Object Relations and Action},
booktitle = {34th British Machine Vision Conference 2023, {BMVC} 2023, Aberdeen, UK, November 20-24, 2023},
publisher = {BMVA},
year = {2023},
url = {https://papers.bmvc2023.org/0893.pdf}
}
Copyright © 2023
The British Machine Vision Conference is organised by