Advancing deep learning-based crowd counting: weakly-supervised approaches, active learning, and uncertainty quantification

Abstract

Crowd counting, an important task in computer vision, involves estimating the number of individuals in a crowded scene, usually from an image or video. Its significance extends across various domains, including urban planning, public safety monitoring, event management, and crowd control. Accurate crowd counting is instrumental in efficient resource allocation, effective crowd management, and informed decision-making in densely populated areas. It enables the assessment of crowd dynamics, prediction of crowd behavior, and maintenance of public safety during large gatherings or events. Moreover, crowd counting is pivotal in infrastructure planning, transportation management, and emergency response preparedness. By providing insights into crowd size and density, crowd counting empowers authorities and stakeholders to formulate effective strategies for crowd control, optimize urban infrastructure, and mitigate potential risks associated with overcrowding.