Dynamic People Counting from Delay-Doppler Images in Challenging Scenarios: Enhancing Model Performance

This study presents a novel radar-based people counting (PCnt) methodology empowered by deep learning (DL) frameworks. The challenges we face include overfitting due to the model's tendency to extract highly domain-specific features. These challenges arise from limited data and clutter in indoor settings. To tackle this, our radar system operates in both lab and industrial environments. We propose a 2D-CNN approach and explore ways to handle these challenges, focusing on improving accuracy through preprocessing techniques. Additionally, we introduced data augmentation strategies to enhance model robustness and mitigate overfitting. Our experiments show our approach accurately counts people moving along the radar line in various environments. However, detecting stationary individuals and distinguishing between moving human and non-human entities remain challenging areas for future work.