Crowd Counting Model Training with the Method of Moments in Electromagnetics

Abstract

We investigate crowd counting for Wi-Fi sensing based on models trained with data from the two-dimensional method of moments for electromagnetic scattering. Those models are applied on measurements with a high number of people. We simulate a crowd with people modelled as dielectric cylindrical shells exposed to an incident electric field from a sensing-enabled Wi-Fi access point. The electric field scattered by the people is computed by the method of moments, taking into account all electromagnetic interactions between the bodies. The scattered field is used to compute channel transfer functions coefficients, from which range and Doppler profiles are derived. A set of channel features, statistical features and impulsive metrics is computed on the profiles, and reduced with Principal Component Analysis. A part of the feature set is used to train a Support Vector Machine classifier, whose output classes are ranges of values of people in the crowd. A good accuracy is reached when testing the classifier on the remaining simulation points. The classifier trained on the method of moments data is then applied on real-life measurements of a crowd, performed with Universal Software Radio Peripherals, and reaches a promising accuracy.