Abstract
This study proposes a lightweight domain-assisted deep learning framework for binary human presence detection using ultra-wideband (UWB) radar. The proposed methodology processes raw UWB radar signals through statistically screened, physics-grounded signal features including Fast Fourier Transform (FFT)-based frequency-domain statistics and Hilbert Transform (HT)-derived envelope statistics which are selected via a per-subject Cohen’s d screening step and stacked as auxiliary input channels alongside the raw signal for a lightweight two-dimensional convolutional neural network (2D-CNN). A cross-subject evaluation protocol (train-on-one-subject, test-on-the-other) is adopted to assess generalization across individuals rather than relying on a pooled, sample-level split. Among the candidate features, a Frequency Standard Deviation (FSTD) is shown to match or exceed the performance of every multi-feature combination tested, indicating that targeted feature selection is more consequential than input fusion for this task. To further improve deployment efficiency, post-training INT8 quantization is applied, reducing the model to approximately 23 KB while preserving classification performance for quantization-robust configurations. Hardware-in-the-loop benchmarking on the STEdgeAI platform indicates on-device inference times ranging from approximately 0.88 ms on AI-enabled STM32N6 hardware to 117–130 ms on STM32H7-class microcontrollers; these figures reflect model inference only and exclude radar acquisition and preprocessing time. Experiments are conducted on a two-subject (one male, one female) indoor dataset; the reported cross-subject results are presented as a relative comparison across feature and quantization configurations rather than as an estimate of population-level generalization. The findings nonetheless illustrate the feasibility of combining principled feature selection with quantization-aware, hardware-validated deployment on embedded artificial intelligence (AI) platforms.