EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

Leakage-Safe Re-Evaluation of Deep Video Models for Violence Detection in Smart-City CCTV Surveillance

Bareen Haval, Sarkar Hasan Ahmed

Dasinya Journal for Engineering and Informatics · 20.08.2026 · 2 (3)

DOI: 10.65542/djei.v2i3.82 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Dasinya Journal for Engineering and Informatics, online erschienen am 20.08.2026. Verfasst von Bareen Haval und 1 weiteren Autor:innen. Die Arbeit nennt 38 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Urban CCTV networks produce more video than human operators can watch, so automatic detection of physical violence has become an active research goal whose progress is measured on shared benchmark datasets; the integrity of those benchmarks is therefore as important as the models tested on them. This study audits the Smart-City CCTV Violence Detection (SCVD) dataset and re-evaluates two deep video models on a corrected version of it. The contribution is methodological: a leakage-auditing and benchmark-reconstruction procedure for video datasets, not a new detection architecture; the two models are evaluated to measure what the leakage conceals. The official train/test division leaks: 287 test clips are byte-for-byte copies of training clips, and a model trained on this division scored a meaningless F1 of 1.000. The benchmark was rebuilt in two stages, first removing exact duplicates with a cryptographic hash and then grouping near-identical clips with a perceptual hash and connected-component analysis, so that clips from one scene cannot fall on both sides of the split. On the corrected benchmark, an efficient three-dimensional network (X3D-S, pre-trained on Kinetics-400) reached a mean F1 of 0.894 (±0.020) and was far steadier across random seeds than a ResNet18+LSTM baseline, whose F1 ranged from 0.751 to 0.894. These findings show that published SCVD scores are inflated by leakage and provide an honest, reproducible baseline for real-time urban violence detection.

Zitieren

APA

Haval, B., & Ahmed, S. H. (2026). Leakage-Safe Re-Evaluation of Deep Video Models for Violence Detection in Smart-City CCTV Surveillance. Dasinya Journal for Engineering and Informatics, 2(3). https://doi.org/10.65542/djei.v2i3.82

Vancouver

Haval B, Ahmed SH. Leakage-Safe Re-Evaluation of Deep Video Models for Violence Detection in Smart-City CCTV Surveillance. Dasinya Journal for Engineering and Informatics. 2026;2(3). doi:10.65542/djei.v2i3.82

BibTeX
@article{Haval2026,
  author = {Haval, Bareen and Ahmed, Sarkar Hasan},
  title = {Leakage-Safe Re-Evaluation of Deep Video Models for Violence Detection in Smart-City CCTV Surveillance},
  journal = {Dasinya Journal for Engineering and Informatics},
  year = {2026},
  volume = {2},
  number = {3},
  doi = {10.65542/djei.v2i3.82}
}
Kostenlos starten →