We are delighted to share some excellent news: Martin Haváček, a doctoral student at our faculty, and his team have succeeded with an extensive review study in Artificial Intelligence Review (Springer). The journal ranks among the very best in its field worldwide: its five-year impact factor is 19.1, it sits in the first quartile (Q1) and the first decile of its category, and it holds an Article Influence Score of 3.567. This is the highest-ranked journal in which a publication related to our faculty has ever appeared (based on metrics as of the article's publication date).
The study, titled Deep Learning for Security-Relevant Event Detection in Visual Data, maps how artificial intelligence can recognise security-relevant situations in camera footage today and where the limits lie that future research still needs to push.
Where safety engineering meets artificial intelligence
The paper grew out of the intersection of two faculties of our university the Faculty of Safety Engineering (FBI) and the Faculty of Electrical Engineering and Computer Science (FEI). Safety engineering brings the question of how to reliably recognise risk situations in CCTV footage; artificial intelligence and computer vision offer the tools to do so automatically and in real time. According to the authors, it is precisely this interdisciplinary combination that allowed the study to succeed in a journal where the world's leading AI groups publish.
The study critically surveys the state of the art across key security tasks from the detection of weapons, violence and anomalous behaviour, through perimeter monitoring, to drone-based surveillance. It focuses on what determines a system's usability in practice: modern deep learning architectures, realistic data collection and annotation (and the fight against dataset bias), robustness in uncontrolled conditions, minimising false alarms, the vulnerability of AI models themselves to adversarial attacks, and the ethical questions of privacy protection.
It is not just whether a system spots a threat — but how fast
Most research to date evaluates detection systems by accuracy: what percentage of weapons, fights or anomalies a model correctly recognises. In real security operations, however, a different quantity often decides the outcome time to detection. A system that catches an event a few seconds earlier gives operators and responders a crucial head start; conversely, even the most accurate model is useless if the warning arrives too late.
This is where the study's original contribution comes in: the Time-to-Detection (TTD) framework. It decomposes the entire chain from the moment an event occurs to the moment an alarm is raised — into individual measurable components, and offers a unified way to compare different approaches precisely in terms of speed. In the process, the authors made a surprising finding: across the analysed literature, no study had systematically tracked time to detection in this way attention has focused almost exclusively on accuracy. TTD is therefore a conceptual framework and a challenge for future research: it offers the vocabulary and the metric with which detection speed could be honestly measured and compared.
Born during a doctoral internship in Latvia
A substantial part of the work was written during Martin's doctoral internship at [name of the Latvian university] in Latvia, whose goal was to prepare a joint publication of the two universities. That goal has been met the resulting study is a joint output of the Czech and Latvian institutions and a fine example of what doctoral mobility can bring.
Alongside Martin Haváček as the lead author, the paper was co-authored by Marek Hutter, Karlis Apalups, Radomír Ščurek a Jan Kubíček, who brought expertise in deep learning, review methodology and the security domain.
Acknowledgements
The research was supported by the LERCO – Life Environment Research Center Ostrava project (reg. no. CZ.10.03.01/00/22_003/0000003) funded by the Just Transition Operational Programme. The authors are grateful for the facilities and support provided.
Citation: Havacek, M., Hutter, M., Apalups, K. et al. Deep learning for security-relevant event detection in visual data: a structured narrative review of the state of the art and future challenges. Artif Intell Rev (2026). https://doi.org/10.1007/s10462-026-11653-z