Real-to-Sim-to-Real Geometry-Aware Fall Trajectory Augmentation for Spatially Robust UWB Fall Detection

Distance-based UWB fall detection is strongly affected by anchor-relative geometry: the same physical motion can produce substantially different range observations when performed at unseen locations or orientations. We address this problem through a real-to-sim-to-real geometry-aware trajectory augmentation framework. Measured human falls are reconstructed as three-dimensional physical trajectories, repositioned and reoriented across the sensing area, and mapped through a geometry-based UWB forward model to generate physically consistent multianchor observations. A representative subset of the generated observations is selected in a learned latent space and used to augment classifier training. Evaluation is performed exclusively on real UWB measurements, with strict spatial separation serving as the principal robustness test. Geometry-aware augmentation improves robustness across Raw, Delta, and Raw+Delta representations for both CNN--GRU and PatchTST classifiers. Moreover, classifiers trained only on generated observations transfer directly to spatially held-out real measurements. These results suggest that a physically grounded observation model can generate useful simulated sensing data: measured trajectories define physical motion, sensor geometry defines the sensing configuration, and the ranging function maps augmented trajectories to UWB observations. The resulting real-to-sim-to-real pipeline expands the spatial support of limited real data without requiring video, reconstructed positions, or anchor coordinates at inference time.