Language-driven Visual Detection in Law Enforcement Video Analysis Systems

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Law enforcement agencies depend on automated video analysis to process large volumes of footage from CCTV networks, body-worn cameras, and drone feeds. Existing AI systems detect and recognise objects drawn from fixed, predefined vocabularies established at training time. An analyst seeking footage of a person carrying a blue umbrella, or a vehicle transporting construction materials, cannot express such queries to a system trained on generic categories. Extending these systems requires new labelled datasets and costly retraining. This paper argues that open-vocabulary object detection and language-driven visual grounding, enabled by recent vision-language models, remove this constraint and represent a paradigm shift for investigative video analysis. We examine the technical foundations, identify operational and governance challenges under the EU AI Act, and propose a research agenda for the information systems community.