EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

A Comparative Analysis of Contrastive and Generative Vision-Language Models for Zero-Shot Behavior Recognition in Surveillance Videos

Ayman Mohamed, Saeed Hamouda, Abdelrahman Elsayed, Mohamed M. Reda Ali

Engineering Systems and Intelligent Technologies (ESIT) · 29.06.2026 · 3 (1) · 23-37

DOI: 10.66279/vcth9h10 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Engineering Systems and Intelligent Technologies (ESIT), online erschienen am 29.06.2026. Verfasst von Ayman Mohamed und 3 weiteren Autor:innen. Die Arbeit nennt 19 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Vision-language models (VLMs) have recently demonstrated strong zero-shot capability for object recognition and scene classification, yet their suitability for modeling covert human behaviors, such as theft, remains largely unexamined. This paper presents a case study comparing two zero-shot VLM paradigms for cashier theft detection in surveillance footage: a contrastive embedding model (CLIP) and a generative vision-language model (a Llama-3.2-11B-Vision-based pipeline operating in the BLIP/BLIP-2 family of generative architectures). On a set of real cashier-counter recordings, the contrastive model produced near-tied confidence scores between theft and normal-activity prompts (theft confidence 0.504, normal-activity confidence 0.496), indicating weak discriminative margin when intent and temporal context are required. The generative pipeline, in contrast, produced confident and structured binary outcomes (theft confidence 1.000, normal-activity confidence 0.000) accompanied by interpretable natural-language descriptions of the suspect and the event. These results, while drawn from a small, non-benchmarked sample rather than a large annotated corpus, suggest that contrastive similarity scoring is better suited to fast object-level screening, whereas generative reasoning is better suited to behavior-level interpretation. A hybrid pipeline that couples a fast contrastive pre-filter with a generative reasoning stage is proposed as a practical direction for zero-shot surveillance systems that require both efficiency and interpretability

Zitieren

APA

Mohamed, A., Hamouda, S., Elsayed, A., & Ali, M. M. R. (2026). A Comparative Analysis of Contrastive and Generative Vision-Language Models for Zero-Shot Behavior Recognition in Surveillance Videos. Engineering Systems and Intelligent Technologies (ESIT), 3(1), 23-37. https://doi.org/10.66279/vcth9h10

Vancouver

Mohamed A, Hamouda S, Elsayed A, Ali MMR. A Comparative Analysis of Contrastive and Generative Vision-Language Models for Zero-Shot Behavior Recognition in Surveillance Videos. Engineering Systems and Intelligent Technologies (ESIT). 2026;3(1):23-37. doi:10.66279/vcth9h10

BibTeX
@article{Mohamed2026,
  author = {Mohamed, Ayman and Hamouda, Saeed and Elsayed, Abdelrahman and Ali, Mohamed M. Reda},
  title = {A Comparative Analysis of Contrastive and Generative Vision-Language Models for Zero-Shot Behavior Recognition in Surveillance Videos},
  journal = {Engineering Systems and Intelligent Technologies (ESIT)},
  year = {2026},
  volume = {3},
  number = {1},
  pages = {23--37},
  doi = {10.66279/vcth9h10}
}
Kostenlos starten →