Improvement of Spatial Robustness in UWB Fall Detection via Temporal Differencing

This letter presents a privacy-preserving fall detection framework based solely on ultra-wideband range sequences. To evaluate robustness beyond conventional same-domain test settings, we introduce a strict spatial split in which training and test windows are separated according to fall regions, so that the test set contains falls from previously unseen parts of the monitored area. We compare the raw distance sequence and its first and second order temporal-difference representations under the same convolutional neural network-gated recurrent unit (CNN-GRU) backbone. To reduce the effect of run-to-run variability, all experiments are repeated 30 times with different random seeds and reported using mean and standard deviation. Results show that, under the original split, the raw distance representation and first-order temporal differencing perform similarly, whereas under the strict spatial split, first-order temporal differencing is clearly superior and substantially more stable across runs. In particular, under the strict spatial split, Δdist achieves the best overall performance with 0.9422 ± 0.0206 accuracy and 0.8342± 0.0290 macro-F1, while the raw dist representation drops to 0.6898±0.1203 accuracy and 0.5769± 0.0977 macro-F1. These findings show that reducing location-dependent bias while preserving motion-related temporal changes is important for spatially robust UWB-based fall detection.