Agentic Multimodal Violence Detection and Human-Gated Emergency Dispatch
International Journal of Advanced Computer Science and Applications · 2026 · 17 (10)
DOI: 10.14569/ijacsa.2026.0171033 ↗Kurz gefasst
Fachartikel in der Zeitschrift International Journal of Advanced Computer Science and Applications, 2026. Verfasst von ASMAE BAALA und 2 weiteren Autor:innen.
Abstract (Originalsprache)
Smart cities deploy dense camera networks, yet automated video analysis usually stops at a detection label, leaving the choice of responder to an overloaded human operator. This study presents an end-to-end multi-agent architecture that carries a violence label from ingestion to an audited, operator-confirmed dispatch decision. Two analyst agents (built with the Agno framework and Google Gemini 3.7 Flash) examine the video and the audio of a surveillance window, each returning a typed threat assessment. A deterministic, rule-based function, with no language model involved, fuses the two verdicts through an entertainment gate and a cross-modal conflict check. Flagged verdicts wait for operator confirmation, routing to police, ambulance, or fire follows fixed rules, and every decision is logged. On a balanced 200-clip XD-Violence subset, fused recall is 0.95 (95% confidence interval 0.89-0.98) at specificity 0.85, with no significant gain over either modality alone or over a single multimodal call, and the first routed responder is a required agency for 70 of the 95 detected violent clips. These figures use a disclosed prompt that takes each clip at face value. Under the production prompt, which tells the analysts to ignore staged violence, both analysts judged every clip to be staged entertainment, such as a film scene, so not one violent clip was flagged (fused recall 0.00). A silent failure in the model’s client library, which turned rate-limit errors into “no threat” verdicts (19 of 30 clips in an earlier run), is characterized and fixed. The integration is feasible and auditable; real surveillance footage remains untested.
Zitieren
APA
BAALA, A., Hanoune, M., & Bentaib, M. (2026). Agentic Multimodal Violence Detection and Human-Gated Emergency Dispatch. International Journal of Advanced Computer Science and Applications, 17(10). https://doi.org/10.14569/ijacsa.2026.0171033
Vancouver
BAALA A, Hanoune M, Bentaib M. Agentic Multimodal Violence Detection and Human-Gated Emergency Dispatch. International Journal of Advanced Computer Science and Applications. 2026;17(10). doi:10.14569/ijacsa.2026.0171033
BibTeX
@article{BAALA2026,
author = {BAALA, ASMAE and Hanoune, Mostafa and Bentaib, Mohssine},
title = {Agentic Multimodal Violence Detection and Human-Gated Emergency Dispatch},
journal = {International Journal of Advanced Computer Science and Applications},
year = {2026},
volume = {17},
number = {10},
doi = {10.14569/ijacsa.2026.0171033}
}
EUVIMED