Fitting device-free crowd size estimation models using area-ratio scaled synthetic labels

Abstract As urbanization accelerates, the ability to accurately monitor crowd densities in public spaces has become essential for enhancing city efficiency, safety, and sustainability. Collecting reliable ground truth data on people counts in large-scale urban environments can be challenging. In this paper, we evaluate an experimental setup that uses device-free wireless sensing to estimate the number of people in an area by analyzing the attenuation of the communication link between the battery-powered sensor nodes of a Wireless Sensor Network (WSN). In this work, we propose a technique that synthesizes count labels based on manual ground truth counts in a smaller portion of the environment to fit a model for the entire environment, thereby simplifying the ground truth gathering process. When comparing the Mean Absolute Percentage Errors (MAPE) for the transformed “subzone-to-total” models, we observe a mean increase of 4.49 % compared to evaluations of models directly fitted on data from the total zone. When transforming between adjacent subzones, we observe a mean increase of 9.65 % compared to evaluations of models fitted directly on the target subzone’s ground truth data. When collecting ground truth data for the total zone is too difficult, labels synthesized via area-ratio scaling from one or more subzones can provide a viable alternative for fitting the estimation model.