Radar-Based Fall Detection for Assisted Living: A Digital-Twin Representation Case Study

TL;DR

This study addresses the scarcity of real-world fall data by leveraging a digital-twin FMCW radar testbed to compare input representations for fall versus non-fall discrimination. From a single simulated range–Doppler sequence, the authors generate three input modes (spec, spec3, and rdm) and evaluate them with an identical compact CNN, revealing that temporal spectrograms provide superior and more balanced performance (near 99% accuracy) than a static time-pooled RDM (~89.4% accuracy). The contributions include a balanced synthetic dataset in a furnished room, a controlled representation comparison, and a qualitative check against measured data to assess realism. The results offer representation-level guidance for radar-based fall detectors within controlled synthetic conditions, while acknowledging the limitations of idealized noise-free twins and single-room configurations for real-world deployment.

Abstract

Obtaining data on high-impact falls from older adults is ethically difficult, yet these rare events cause many fall-related health problems. As a result, most radar-based fall detectors are trained on staged falls from young volunteers, and representation choices are rarely tested against the radar signals from dangerous falls. This paper uses a frequency-modulated continuous-wave (FMCW) radar digital twin as a single simulated room testbed to study how representation choice affects fall/non-fall discrimination. From the same simulated range-Doppler sequence, Doppler-time spectrograms, three-channel per-receiver spectrogram stacks, and time-pooled range-Doppler maps (RDMs) are derived and fed to an identical compact CNN under matched training on a balanced fall/non-fall dataset. In this twin, temporal spectrograms reach 98-99% test accuracy with similar precision and recall for both classes, while static RDMs reach 89.4% and show more variable training despite using the same backbone. A qualitative comparison between synthetic and measured fall spectrograms suggests that the twin captures gross Doppler-time structure, but amplitude histograms reveal differences in the distributions of amplitude values consistent with receiver processing not modeled in the twin. Because the twin omits noise and hardware impairments and is only qualitatively compared to a single measured example, these results provide representation-level guidance under controlled synthetic conditions rather than ready-to-use clinical performance in real settings.

Figures (8)

• Figure 1: Discrete grid over subject position and yaw relative to the radar, with one example fall trajectory shown.
• Figure 2: Digital-twin simulation setup. (a) Top-down and side-view schematic of the $6 \times 8~\text{m}$ furnished room, showing the ceiling-mounted FMCW radar, its field of view, and example reflecting objects in the modeled environment. (b) Perceive-EM renderings of one fall at three time instants.
• Figure 3: Empirical amplitude histograms (in dB) over all pixels for the three per-RX Doppler--time spectrograms. The means are similar but not identical, suggesting that each receive patch sees a slightly different multipath power distribution rather than an exact copy of the same signal.
• Figure 4: Example Doppler--time spectrograms at the fixed range bin $r^\star \approx 3.9~\text{m}$ for a non-fall (a) and a fall (b) clip in the digital twin.
• Figure 5: Test accuracy versus epoch for each representation mode. The spectrogram modes approach and stay near $99\%$, while the time-pooled RDM mode follows a lower curve with larger fluctuations over epochs.