Inception-ResNet 1-D CNN With Subject’s Physical Information for Reliable Real-Time FMCW Radar Fall Detection

Reliable real-time fall detection using radar sensing remains challenging due to the strong similarity between fall events and fall-like daily activities such as squatting, bending or sitting, as well as the sensitivity of deep learning models to environmental variations during real-time deployment, such as changes in sensing conditions. In particular, false alarms frequently arise near the edges of the region of interest (ROI), reducing the stability of deep learning–based radar systems in practical deployments. To solve this issue, this study proposes a privacy-preserving fall detection system based on frequency-modulated continuous-wave (FMCW) millimeter-wave (mmWave) radar by integrating theInception–ResNet1D architecture with a physics-informed decision layer calledSubject’s Physical Information(SPI). The proposed system combines deep learning–based temporal modeling with lightweight physical validation derived from vertical motion cues, including the minimum height (zmin), vertical displacement (Δz), and downward velocity ratio (rv−z), to stabilize real-time decisions. The system is implemented in a multithreaded architecture with bounded real-time latency. The learning model was trained using data collected from ten subjects, while experiments were conducted on five different subjects distinct from the training cohort for real-time evaluation within a 3 × 3 m area covering both center and edge ROI. Compared with the baselineInception–ResNetmodel without SPI, the proposed system improves precision from 62.71% to 84.90%, accuracy from 70.27% to 91.11%, and F1-score from 77.09% to 91.84%, while maintaining a recall of 100% from tests on five subjects under controlled room conditions. In addition, a supplementary robustness evaluation in an uncontrolled-room setting further shows that SPI integration helps maintain improved overall system performance under more challenging sensing conditions. These results indicate that combining deep learning-based prediction with lightweight physical validation yields a more reliable and stable radar-based fall detection system suitable for real-time deployment in elderly monitoring applications.