EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

STORM-Net: Summary-Token Object-Relational Mamba Network for Dynamic Crime-Scene Graph Generation

Rajeshwar Yadav, Raju Halder, Gourinath Banda

IEEE Access · 2026 · 14 · 98132-98147

DOI: 10.1109/access.2026.3707528 ↗

Kurz gefasst

Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Rajeshwar Yadav und 2 weiteren Autor:innen. Die Arbeit nennt 64 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

The recognition of crime incidents in surveillance videos is a critical challenge for public safety. Yet, existing approaches remain limited in capturing the complexity of real-world crime events. Current methods often operate at a coarse video level, i.e., they produce binary predictions without revealing crucial details about victims, offenders, weapons, or evolving interactions. Dynamic Scene Graph Generation (DSGG) offers a promising paradigm to address these limitations. It represents videos as structured graphs of objects and their relationships, and enables interpretable reasoning about “who” is interacting with “whom” and “how” these interactions evolve over time. However, existing DSGG models are often restricted to short temporal windows. As a result, they fail to capture long-range dependencies that are essential for accurate prediction, where critical cues may be dispersed across non-contiguous frames. To overcome this limitation, we propose STORM-Net (Summary Tokens with Object-aware Relational Mamba Network), a DSGG framework designed to model extended spatio-temporal dependencies. STORM-Net introduces a learnable summary token mechanism. It compresses local frame sequences into compact representations for scalable global reasoning. In parallel, Mamba layers are integrated into both the encoder and decoder. They recurrently capture evolving object relations across extended time horizons. Furthermore, we present Relational Scene Understanding in Complex Hostage Events (RESCUE). RESCUE is a dataset for crime-based group activity recognition, consisting of detailed multi-person annotations in realistic hostage and threat scenarios. Experimental evaluations on Action Genome and RESCUE demonstrate that STORM-Net consistently outperforms state-of-the-art baselines, establishing a robust and interpretable foundation for crime-scene analysis.

Zitieren

APA

Yadav, R., Halder, R., & Banda, G. (2026). STORM-Net: Summary-Token Object-Relational Mamba Network for Dynamic Crime-Scene Graph Generation. IEEE Access, 14, 98132-98147. https://doi.org/10.1109/access.2026.3707528

Vancouver

Yadav R, Halder R, Banda G. STORM-Net: Summary-Token Object-Relational Mamba Network for Dynamic Crime-Scene Graph Generation. IEEE Access. 2026;14:98132-98147. doi:10.1109/access.2026.3707528

BibTeX
@article{Yadav2026,
  author = {Yadav, Rajeshwar and Halder, Raju and Banda, Gourinath},
  title = {STORM-Net: Summary-Token Object-Relational Mamba Network for Dynamic Crime-Scene Graph Generation},
  journal = {IEEE Access},
  year = {2026},
  volume = {14},
  pages = {98132--98147},
  doi = {10.1109/access.2026.3707528}
}
Kostenlos starten →