Large-Scale Crowd Size Classification With a Wi-Fi-Based Passive Radar

We investigate the problem of large-scale crowd size classification with a Wi-Fi-based passive radar for crowds of up to 100 people, with classes corresponding to intervals of numbers of people. A convolutional neural network (CNN) operating on radar range-Doppler maps (RDMs) is used as a classification algorithm. We propose a crowd simulator based on the method of moments (MoM) in electromagnetics able to generate representative RDMs for large crowds. We show that these MoM simulation data can be used to design the classification algorithms and tune their hyperparameters. We also investigate the limitations of the MoM simulation data in training the classification algorithms for subsequent application on experimental data. Crowd size classification is performed with high accuracy on real-life experimental measurements of a crowd with up to 100 people, obtained by channel estimation with 802.11ax-compliant high-efficiency long training fields (HE-LTF) transmitted by a Wi-Fi-based passive radar setup featuring two Universal Software Radio Peripherals (USRPs) X310.