Answers are grounded in this paper and cite their sources.
Crowd counting refers to the task of estimating the number of people within a specific area, which provides insights into crowd dynamics and distribution. This task has widespread applications across various domains, including public safety, urban planning, and traffic management. Recent advances in deep learning, particularly the convolutional neural networks and the Transformer architecture, as well as multimodal pre-trained models, such as CLIP and SAM, have substantially improved counting accuracy. However, existing models are mainly designed in high-complexity architecture, which incur heavy computational and data requirements, and hinder real-time deployment and increase the cost. Consequently, lightweight crowd-counting approaches have emerged as a key research direction, which aims to reduce parameters and inference time while retaining high performance. In this survey, we summarize publicly available datasets, evaluation metrics, latest lightweight architectures, and evaluate representative models on benchmark datasets to inspire future research in crowd counting.
No takes yet. Share an insight, caveat, or question.
Zhai et al. (2025) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: