A system for detecting deviant behavior based on facial video sequences in public transport

S.A. Onopko, J.A. Pavelina, J.I. Minkin, P.A. Popov

Abstract


This paper examines the problem of detecting deviant behavior in facial video sequences. Unlike static analysis of individual frames, the study relies on temporal patterns of facial expressions and physiological states observed throughout the video. To test the hypothesis that vision-language models are applicable to video interpretation, the gpt5.5-quick model was used, and the evaluation results were recorded using two groups of features: emotions and states. It is demonstrated that the system has sufficient accuracy for solving applied problems of automated emotional state analysis and detecting signs of deviant behavior in video data. The results confirm that the approach works at the proof-of-concept level; the next stage involves local deployment of an open vision-language model on more powerful in-house hardware.


Full Text:

PDF (Russian)

References


Jeganathan J. et al. Quantifying dynamic facial expressions under naturalistic conditions //Elife. – 2022. – Т. 11. – С. e79581.

Zhao J. et al. Facial dynamics in video: Instruction tuning for improved facial expression perception and contextual awareness //Proceedings of the AAAI Conference on Artificial Intelligence. – 2026. – Т. 40. – №. 16. – С. 13163-13171.

Cuve H. C. J., Sowden-Carvalho S., Cook J. Spatiotemporal structure and substates in emotional facial expressions //bioRxiv. – 2024. – С. 2024.09. 02.610622.

Wu L. et al. RoboStressBench: Benchmarking VLM Robustness to Physical Visual Stress in Embodied Scenes //arXiv preprint arXiv:2606.00828. – 2026.


Refbacks

  • There are currently no refbacks.


Abava  Кибербезопасность ИТ-КОНГРЕСС ВМК МГУ 2026 СНЭ

ISSN: 2307-8162