Lightweight Edge Model for Real-Time Breathing-Based Human Presence Detection with UWB Radar
27th International Radar Symposium, IRS 2026, Krakow, Poland, 19 - 21 May 2026, pp.15-20, (Full Text)
- Publication Type: Conference Paper / Full Text
- Doi Number: 10.23919/irs70539.2026.11548971
- City: Krakow
- Country: Poland
- Page Numbers: pp.15-20
- Keywords: Convolutional Neural Network, Edge-Deployment, Human Detection, UWB Radar
- Istanbul Kültür University Affiliated: No
Abstract
This study presents an edge-deployable Convolutional Neural Network (CNN) model designed for fast and efficient human presence detection using raw ultra-wideband (UWB) radar data. The objective is to minimize decision time and enable deployment on edge devices with limited computational resources through model quantization techniques. Data were collected from two human subjects positioned at various distances and angles relative to the radar. To enhance feature representation, frequency-domain and amplitude-based analyses were applied using the Fast Fourier Transform (FFT), Root Mean Square (RMS), and Hilbert envelope methods. In the final step, Occlusion Sensitivity and Integrated Gradients have been applied to reveal the temporal and spatial patterns that drive the model's human detection decisions.