Hybrid DCNN–Transfer Learning Model Coupled With Background Clutter Mitigation for FMCW Radar-Based People Counting Improvement

Abstract

Automatic people counting has garnered significant attention due to its broad civilian and military applications. In civilian settings, it helps detect unusual occupancy patterns or manage crowding in public transportation. In military contexts, it serves to count and track enemy movements, providing real-time data on troop numbers and positions on the battlefield, which is critical for tactical decision-making. Radar systems are often used for such tasks due to their ability to function in all weather conditions, day or night. However, the signal collected by the radar is hindered by unwanted signals reflected by clutter. Also, the direct coupling between transmit and receive antennas can mask targets with a weak signal. All these artifacts can decay the performance of deep learning models for automatic people counting. This work proposes a background mitigation algorithm based on the multiresolution analysis of the maximal overlap discrete wavelet transform (MRA-MODWT) to enhance the accuracy of deep learning models for automatic people counting. Subsequently, the Daubechies least asymmetric wavelet with four vanishing moments (sym4) is used to isolate and cancel background signals, and a fusion layer combining a transfer learning block with a customized deep convolutional neural network (DCNN) is introduced to improve the accuracy. The hybrid DCNN-InceptionV3 model achieved a peak accuracy of 98.31%, an average precision of 0.9827, an average recall of 0.9827, and an average $F1$ score of 0.9836 on realistic radar data.