FMCW Radar Crowd Counting with Beamforming and Attention-Enhanced Multi-Scale CNN

In this research, we offer a novel approach for crowd counting employing 77-GHz frequency-modulated continuous wave (FMCW) radar, combined with digital beamforming and a multiscale convolutional neural network (CNN) improved by a channel attention mechanism. The system generates 2D spectral waterfall heat maps as inputs to multiple CNN channels using beamforming to capture radar data from various perspectives. The channel attention mechanism prioritizes signals from critical directions, improving detection accuracy. Experimental validation of data obtained from up to eight persons demonstrates that the proposed method achieves a high accuracy of 98.2 %, outper-forming models without beamforming and attention processes.