A 3D CNN based People Counting System Using Auto-Correlation Functions from Frequency Modulated Continuous Wave Radar Signals

We propose a people counting system using Frequency Modulated Continuous Wave (FMCW) radar signals. The proposed method predicts the number of people using deep learning classification with auto-correlation functions of estimated breathing signals. We also attempted to explain the behavior of the deep learning system using modified gradient-weighted class activation mapping (GradCAM). In the experiments using real radar signals, the proposed method showed improved performance from a conventional method.