Abstract

Crowd monitoring is essential for smart city applications, particularly for optimizing public transit systems. To address this need, we propose a privacy-conscious crowd-counting pipeline using Wi-Fi probe requests. This pipeline is designed to adapt to the challenges posed by the randomization of Media Access Control (MAC) addresses, which serve as unique identifiers for devices on a network. Our approach leverages a random forest-based feature selection process to identify key Information Elements and frame attributes then applies DBSCAN clustering with adaptive parameter optimization for device counting. A diffusion model generates synthetic tabular data to mitigate the limited availability of labelled data, enhancing model robustness. Experimental results demonstrate improved accuracy in device counting, achieving a V-measure of 0.952, an average silhouette score of 0.789, and reliable clustering counts.