Fall detection no longer needs the full IMU corpus: imbalanced recordings compress to dozens of synthetic windows, and a tiny network still clears ninety percent
Synopsis
Training data for wearable fall detection is compressed into a very small class-balanced synthetic set. Earlier pipelines either used hand-crafted features with shallow classifiers or trained deep networks on the full imbalanced corpus with re-sampling, at high data and compute cost. FallNest aligns synthetic and real samples in both time and frequency domains through gradient direction and embedding distribution, initialized by class-balanced K-Means and augmented with multi-view Fourier views. On SisFall, KFall and MobiFall under subject-wise splits with naturally imbalanced test sets, a 102K-parameter TinyCNN trained on only 32 to 146 synthetic windows reaches 93.5 to 95.8 percent mean accuracy and 0.89 to 0.98 PR-AUC over five seeds.
Interpretation
The training data a fall detector needs shrinks to roughly one percent, and class balance comes from the condensed set itself rather than from re-sampling. Earlier pipelines either used hand-crafted features with shallow classifiers, which limits learning from raw signals, or trained deep networks on the full imbalanced corpus, which needs re-sampling and substantial data and compute. On SisFall, KFall and MobiFall under subject-wise splits with naturally imbalanced test sets, a 102K-parameter TinyCNN trained on only 32 to 146 synthetic windows reaches 93.5 to 95.8 percent mean accuracy and 0.89 to 0.98 PR-AUC over five seeds, with accuracy standard deviations of 1.0 to 1.7 points.
Synthetic samples are pulled toward real data in two spaces at once, time and frequency, so a handful of samples still covers the main modes of both classes. CondTSC condensed time-series data into synthetic samples but was not evaluated on wearable fall detection under subject-wise splits with naturally imbalanced test sets. Each sample is seen through the raw signal and three Fourier-domain views, gradient matching uses a normalized squared distance with gradient clipping, embedding alignment uses Gaussian-kernel MMD, and initialization uses class-wise K-Means, with all hyperparameters fixed across datasets.
On the two datasets with the smallest synthetic sets, condensed data beats real subsets of the same size. Equal-size real subsets are drawn either uniformly or with equal numbers per class, under the same TinyCNN, checkpoint selection and five-seed protocol. On MobiFall and KFall, FallNest reaches PR-AUC 0.975 and 0.973 while random subsets reach at most 0.932 and 0.90; on SisFall the uniform subset is slightly ahead in accuracy, 94.0 versus 93.5, so the advantage over uniform sampling there is not established.
Perspective
The results cover three public datasets, SisFall, KFall and MobiFall, under subject-wise splits with naturally imbalanced test sets, reporting accuracy, Macro-F1 and PR-AUC over five seeds. The intended readers are researchers and engineers who must train wearable fall classifiers under limited storage or compute. Inference latency, memory and energy follow from the architecture and are not changed by condensation; microcontroller validation and false-alarm analysis on free-living data are listed as future work.
Against full-data training, accuracy drops by 3.6 to 6.4 points, and SisFall PR-AUC falls from 0.999 to 0.891. On SisFall the uniform real subset reaches 94.0 accuracy versus 93.5, so the advantage over uniform sampling is not established. The ablations do not isolate the accelerometer vector magnitude, normalized versus cosine matching, or clipping, and no CondTSC or other coreset baseline is included. Subject splits are fixed across seeds, so reported standard deviations reflect initialization variability only. The constants in the theoretical bounds are unmeasured, and the bounds concern the training distribution.