Simulation-to-Real Trip-Fall Detection with Continuous-Wave Doppler Radar via Physics-Informed Kinematic Modeling and Domain Randomization.

Falls among older adults are a major public health concern, yet collecting large-scale real fall data for radar-based detection is ethically and practically difficult. This study presents a controlled simulation-to-real feasibility study for trip-fall detection using continuous-wave (CW) Doppler radar. The method couples a physics-informed kinematic trip-fall model with a CW radar observation model to synthesize I/Q signals and Doppler spectrograms, while domain randomization varies body size, fall direction, initial velocity, sensor placement, aspect angle, amplitude, and noise. Synthetic walking and respiration data were also generated for controlled three-class classification among trip fall, walking, and seated quiet breathing. In Experiment I, the simulated spectrograms reproduced the dominant time-frequency characteristics of measured enacted trip-fall signals acquired with a 24 GHz CW radar; quantitative similarity analysis yielded a mean SSIM of 0.782 and a Doppler-ridge MAE of 24.6 Hz across five fall directions. In Experiment II, a ResNet-18 classifier trained only on simulated spectrograms achieved a macro-F1 score of 0.912 [95% CI: 0.883-0.936] on measured data from ten participants, three start locations, and eight directions. Under the present controlled evaluation, this exceeded the available real-data-trained baseline of 0.748 [95% CI: 0.691-0.805] (paired subject-level permutation test, p=0.006). These findings suggest that physics-informed simulation with domain randomization can reduce dependence on real trip-fall samples under limited-data conditions. The results do not establish robustness to other fall morphologies, fall-like activities of daily living, different environments, different radar devices, or embedded deployment.