A New BLIP based Natural Language Analysis Method for Crowd Images

Abstract

Automatic reporting of crowd density and spatial distribution is a significant problem for smart city and campus applications. This study presents a BLIP-based visual language model approach that can generate density reports in natural language from crowd images. In the proposed method, five synthetic descriptions encompassing the number of people, density level, and spatial distribution information for each image in the UCF-QNRF dataset are generated using a rule-based template structure. This structure provides linguistic diversity by using different sentence patterns that express the same quantitative information. The automatic reporting performance of the BLIP model trained with the generated synthetic descriptions was analyzed using ROUGE and CIDEr metrics, which are evaluation criteria suitable for text generation tasks. Experimental results show that the proposed approach can produce consistent and meaningful descriptions in crowd density reporting; achieving performance values of 47.79 for ROUGE-L and 44.1 for CIDEr. © 2026 IEEE.