Augmented High-Resolution Range Profile for Recurrent Neural Network-Based People Counting Using Millimeter-Wave FMCW MIMO Radar

This study proposes a novel framework to enhance automatic people counting using millimeter-wave frequency-modulated continuous-wave (FMCW) multiple-input–multiple-output (MIMO) radar. The approach leverages an augmented high-resolution range profile (HRRP) as input to a recurrent neural network (RNN). Conventional single-channel HRRP inputs often lead to instability during training and suboptimal performance due to radar-specific limitations, such as false alarm, missed detection, and limited resolution in distinguishing closely spaced targets, particularly in regions where moving target trajectories intersect. To address this limitation, we introduce a multichannel input representation referred to as the augmented HRRP, comprising the in-phase and quadrature (I/Q) components of the time-domain radar signal, as well as the magnitude and phase of the corresponding HRRP. The proposed model is evaluated using a hold-out validation scheme on real-world radar measurements collected across 11 scenarios involving dynamic targets, with the number of individuals varying from 0 to 10. The highest performance is achieved using a hybrid convolutional neural network-gated recurrent unit (CNN-GRU) architecture, yielding a peak accuracy of 95.07%, a macro-averaged precision of 0.9511, a macro-averaged recall of 0.9502, and a macro-averaged$F1$-score of 0.9506.