PaperlayerView exact revision on arXiv

Source-linked AI summary

FALLNEST: Dual-Domain Dataset Condensation of Imbalanced IMU Data for Wearable Fall Detection

Rashi Niyas P, Ritika Jain, Hitika Tiwari

arXiv:2610.18260v1cs.CVcs.AIeess.SP

TL;DR

Wearable fall detection must contend with large, imbalanced IMU datasets and comparisons whose evaluation protocols vary. FALLNEST condenses such data into tiny, balanced synthetic sets; across three datasets, a 102K-parameter TinyCNN trained on 32–146 windows achieved 93.5–95.8% mean accuracy and 0.89–0.98 PR-AUC.

  • Problem

    Fall detection must be learned from large, imbalanced IMU datasets, while differing evaluation protocols complicate comparisons with full-data methods.

    Source: Method · p. 2, Comparison with Full-Data Methods · p. 7

  • Method

    FALLNEST uses dual-domain condensation to create a tiny, class-balanced IMU training set and trains a lightweight classifier on that set alone.

    Source: Method · p. 2

  • Results

    93.5–95.8% mean accuracy and 0.89–0.98 PR-AUC were reported across three datasets for a 102K-parameter TinyCNN trained on 32–146 synthetic windows.

    Source: Conclusion · p. 9

  • Takeaways & Limitations

    FALLNEST offers a compact-data training option evaluated under strict subject-wise splits and naturally imbalanced test sets.

    Source: Conclusion · p. 9

    Ablations do not isolate AVM, normalized versus cosine matching, or clipping, and no CondTSC or other coreset or condensation baselines were evaluated.

    Source: Ablation Studies · p. 8

Abstract

from arXiv · show

Wearable fall detection is hampered by rare fall events, scarce labeled IMU data, and the cost of training deep models on large, imbalanced corpora. Existing methods either rely on hand-crafted features with shallow classifiers, which limits learning from raw signals, or train deep networks on full datasets that require re-sampling to counter class skew and demand substantial data and compute. To address this, we propose a novel Fall detection framework via Normalized-gradient and Embedding matched Synthetic Time-frequency condensation, dubbed FALLNEST, a data-centric approach that distills large, imbalanced IMU datasets into tiny, class-balanced synthetic training sets. Building on dual-domain time-series condensation (CondTSC), FALLNEST introduces a fall-specific 7-channel input with accelerometer vector magnitude, class-balanced K-Means initialization, and normalized gradient matching with gradient clipping, combined with MMD-based embedding alignment and multi-view Fourier-domain augmentation, so that class balance arises from the condensed set itself rather than from re-sampling. We evaluate on SisFall, KFall, and MobiFall using strict subject-wise splits and naturally imbalanced test sets, reporting accuracy, Macro-F1, and PR-AUC over five seeds. A 102K-parameter TinyCNN trained on only 32 to 146 synthetic windows (about only 1% of the training windows) reaches 93.5-95.8% mean accuracy and 0.89-0.98 PR-AUC, with standard deviations of 1.0-1.7% in accuracy, demonstrating the efficacy of our approach.

Method · Data Preparation

FALLNEST condenses imbalanced IMU training data into a tiny, class-balanced synthetic set using K-Means initialization and time- and frequency-domain matching, then trains a TinyCNN on that set.

Method

  • FALLNEST trains the TinyCNN on synthetic set S, selects it on real validation set V, and evaluates it on subject-independent, naturally imbalanced test set Ttest.The real training, validation, and test sets are subject-disjoint.Source: Method · p. 2
  • The synthetic set contains spc = ⌈0.01 n/2⌉ samples per class, for |S| = 2 spc ≈ 0.01 |T| overall.Here n counts all training windows, and each synthetic sample has a fixed class label.Source: Method · p. 2
  • Condensation aims for a network trained only on S to perform on real data like one trained on T, replacing costly exact optimization with matching objectives.The exact objective would require a training run for every update.Source: Method · p. 2

Data Preparation

  • SisFall, KFall, and MobiFall recordings are resampled to 200 Hz and segmented into 10-, 6-, and 5-second windows, respectively.The datasets contain 15, 15, and 4 fall types and 19, 21, and 9 ADL types; all falls map to 1 and ADLs to 0. The 2-second stride and window lengths of 2000, 1200, and 1000 samples were chosen from each dataset’s mean fall duration.Source: Data Preparation · p. 3
  • Windows inherit their source trial’s binary label, including transition windows, and splits are subject-wise with the same split across all five seeds.Subject counts for training, validation, and test are 26/6/6 for SisFall, 16/4/4 for MobiFall, and 15/4/13 for KFall.Source: Data Preparation · p. 3
  • Each window has seven channels: tri-axial accelerometer and gyroscope signals plus accelerometer vector magnitude, with each channel standardized using training-set statistics.No global low-pass filter is applied to avoid smoothing the impact peak; smoothing is used only as an augmentation view.Source: Data Preparation · p. 3
  • Class-wise K-Means clusters flattened windows into spc centroids per class, which are reshaped as synthetic samples so S is balanced regardless of the real class ratio.Centroids cover dominant class modes, particularly within heterogeneous ADL data, but average away sharp peaks that refinement must restore.Source: Data Preparation · p. 3

Dual-Domain Representation

FALLNEST represents each window in both the time and frequency domains, using separate domain-specific backbones shared between synthetic and real data.

  • The time-domain input is the window itself, while the frequency-domain input is a channel-averaged magnitude spectrogram.Source: Dual-Domain Representation · p. 4
  • With nfft = 256, hop 64, and 129 frequency bins, the spectrogram has 28, 15, and 12 frames for SisFall, KFall, and MobiFall, respectively.Each domain has its own backbone, shared between synthetic and real data.Source: Dual-Domain Representation · p. 4

Multi-View Fourier Augmentation · Dual-Objective Matching

FALLNEST augments each sample with raw and Fourier-domain views, then condenses data using a dual objective that aligns gradient behavior and class-wise embeddings. Its algorithm initializes class centroids, optimizes synthetic samples across views and domains, and selects a fresh TinyCNN by validation Macro-F1.

Multi-View Fourier Augmentation

  • Each sample has four views: raw, low-pass at fc = 5 Hz, phase-perturbed with Φ ∼ N(0, 0.12), and magnitude-perturbed with Γ ∼ N(0, 0.1).The low-pass view emphasizes macro-motion; phase and magnitude perturbations target tolerance to small timing shifts and intensity variation, respectively.Source: Multi-View Fourier Augmentation · p. 4

Dual-Objective Matching

  • Normalized squared distance aligns synthetic and real gradients while penalizing both direction and magnitude mismatch.Each class-wise term lies in [0, 2], regardless of dataset size, and is comparable to the bounded MMD term.Source: Dual-Objective Matching · p. 4
  • Penultimate-layer embeddings from synthetic and real samples of the same class are aligned with squared MMD using a Gaussian kernel with bandwidth σ = 1.The passage states that ϵ > 0 is a small constant and each class-wise gradient term is bounded in [0, 2].Source: Dual-Objective Matching · p. 4
  • The total objective combines gradient matching of how the network learns with weighted embedding matching of what it represents.The embedding term is weighted by λemb.Source: Dual-Objective Matching · p. 4
  • The method backpropagates through the unrolled inner loop and clips ∇SL elementwise, then trains a fresh TinyCNN and keeps the epoch with best validation Macro-F1.The clipping threshold is not specified in the supplied passage.Source: Dual-Objective Matching · pp. 4–5
  • The algorithm initializes each class with K-Means centroids, then optimizes synthetic data for 100 epochs across views, domains, and classes using Lgrad + λembLemb.Each loop draws initial parameters, performs N synthetic-data SGD steps and M steps on real batches, then updates S with Adam.Source: Dual-Objective Matching · p. 5

Downstream Training, Backbone, and Evaluation

A fresh time-domain TinyCNN is trained only on the condensed set S, selected by Macro-F1 on real validation data V, and evaluated once on Ttest. Evaluation uses five independent seeds and reports classification metrics, with window-level dependence and the laboratory-data scope of FA/h noted.

Downstream Training, Backbone, and Evaluation

  • The deployed time-domain TinyCNN has 102 K parameters and uses three Conv1D–BatchNorm–ReLU blocks, while the frequency branch is used only during condensation.It is trained on S alone with cross-entropy for 100 epochs using Adam (lr 0.001) and batch size 16; its architecture also includes adaptive average pooling and a fully connected layer.Source: Downstream Training, Backbone, and… · p. 5
  • Each of five seeds independently initializes K-Means, runs condensation, and trains a new classifier on the same subject-wise split; the highest real-validation Macro-F1 checkpoint is evaluated once on Ttest.Condensation uses T alone, so V is the only other real data used.Source: Downstream Training, Backbone, and… · p. 5
  • Reported metrics are mean ± std for Accuracy, Macro-F1, and Fall-positive PR-AUC, plus fall precision, recall, specificity, and false alarms per hour (FA/h).FA/h counts false-positive windows per ADL hour at a 2 s stride and is a window-level estimate on laboratory data; overlapping windows from one trial are correlated, so window counts overstate independent events.Source: Downstream Training, Backbone, and… · p. 5
  • Inference requires about 103 M MACs at L=1000 and 206 M MACs at L=2000, with 2.15 ms median CPU latency at batch size 1.The 2.15 ms latency was measured on Kaggle CPU with single-threaded PyTorch; these are architecture properties, and microcontroller validation remains future work.Source: Downstream Training, Backbone, and… · p. 5

Theoretical Analysis of the Objective

Under stated assumptions, the objective controls gradient-update direction and scale and bounds the real-loss gap; these qualitative bounds concern the training distribution, not unseen subjects.

Theoretical Analysis of the Objective

  • The analysis assumes a Lipschitz, smooth real loss, bounded SGD steps, an RKHS-approximable head loss, and i.i.d. real embeddings conditional on S.The loss approximation has norm bound B and error bound κ.Source: Theoretical Analysis of the Objective · p. 5
  • The corollary bounds gradient alignment by cos ∠(GS, GT) ≥ 1 − δ, the gradient-norm ratio by [ρ−, ρ+], and the real-loss gap by Λ∥∆∥.It also gives a probability bound over the real batch, but the supplied passage omits the bound’s expression.Source: Theoretical Analysis of the Objective · p. 5
  • Eq. (8) controls update direction and scale, while Eq. (9) lets the MMD term bound real loss by synthetic loss; the resulting bounds are qualitative and do not cover unseen subjects.The constants are unmeasured, and the bounds concern the training distribution.Source: Theoretical Analysis of the Objective · p. 6

Experimental Setup · Datasets and Implementation

FALLNEST is evaluated on SisFall, KFall, and MobiFall with subject-wise splits. Condensation settings are largely fixed across datasets, and five fixed seeds measure initialization variability under fixed splits.

Datasets and Implementation

  • The evaluation covers SisFall (38 subjects), KFall (32), and MobiFall (24) for binary fall detection.Dataset statistics and subject-wise splits are summarized in Table 1.Source: Datasets and Implementation · p. 6
  • Condensation uses at most 100 epochs, learning rate 0.01, λemb = 1, N = 15, M = 100, inner learning rate 0.001, patience 10, and fc = 5 Hz.These hyperparameters are fixed across datasets; no dataset-specific tuning is performed.Source: Datasets and Implementation · p. 6
  • The condensation spc values are 73 for SisFall, 16 for MobiFall, and 18 for KFall; real-data batch sizes are 128 for SisFall and 64 for KFall and MobiFall.The real-data batch size is the sole exception to otherwise fixed condensation hyperparameters.Source: Datasets and Implementation · p. 6
  • Models are implemented in PyTorch and trained on a 16 GB P100 GPU, using five fixed random seeds.Seed 42 is used for representative single-run results.Source: Datasets and Implementation · p. 6
  • Because subject splits remain fixed across seeds, reported standard deviations reflect initialization variability rather than variation across subject partitions.The evaluation uses fixed subject-wise splits.Source: Datasets and Implementation · p. 6
  • Table 1 defines T as the condensation set, V as the checkpoint-selection set, and Ttest as a once-used test set with its natural class distribution.The table describes dataset statistics after subject-wise splitting with a 2 s stride.Source: Table 1 · p. 6

Results and Discussion · Classification Performance · Comparison with Equal-Budget Real Subsets

FALLNEST condenses imbalanced IMU training data to 1% while achieving strong subject-independent classification results; equal-budget real-subset comparisons show dataset-dependent benefits.

Classification Performance

  • Across five seeds, accuracy standard deviations are 1.0–1.7 points and PR-AUC standard deviations are 0.013–0.023; fixed test subjects make these initialization variability only.The tests use fixed subjects, so the reported standard deviations do not measure variation across test subjects.Source: Classification Performance · p. 6
  • SisFall’s larger performance gap may relate to lower fall prevalence, longer windows that dilute impact signatures, and greater ADL heterogeneity relative to its 146-sample condensed set; these hypotheses require targeted evaluation.The passage presents these explanations as hypotheses, not established causes.Source: Classification Performance · p. 6
  • With 1% of training windows, FALLNEST reaches 93.5–95.8% accuracy and 0.891–0.975 PR-AUC, exceeding always-ADL accuracy by 9.7, 20.7, and 30.3 points on SisFall, MobiFall, and KFall.Table 2 reports test Accuracy, Macro-F1, and PR-AUC on naturally imbalanced, subject-independent test data.Source: Classification Performance · p. 6, Table 2 · p. 7
  • Compared with full-data training, FALLNEST accuracy is lower by 6.4 points on SisFall, 4.5 on MobiFall, and 3.6 on KFall; SisFall PR-AUC falls from 0.999 to 0.891.Full-data results use a single run, whereas condensed-set results are means over five seeds.Source: Classification Performance · p. 6, Table 2 · p. 7

Comparison with Equal-Budget Real Subsets

  • On MobiFall and KFall, with 32 and 36 samples, FALLNEST leads equal-budget random subsets in accuracy and PR-AUC, reaching 0.975 and 0.973 PR-AUC versus random-subset maxima of 0.932 and 0.90.The comparison uses the same TinyCNN, checkpoint selection, and five-seed protocol; random subsets have the same size as the condensed set and are drawn uniformly or with equal numbers per class.Source: Comparison with Equal-Budget Real… · pp. 6–7, Table 3 · p. 7
  • On SisFall, the uniform subset is competitive in accuracy (94.0 versus 93.5) and Macro-F1, while FALLNEST improves PR-AUC over the class-balanced subset (0.891 versus 0.765); benefit over uniform sampling is not established.The passage attributes the ranking-quality gain to the comparison with the class-balanced subset, not to uniform sampling.Source: Table 3 · p. 7, Comparison with Equal-Budget Real… · p. 7
  • On MobiFall and KFall, the class-balanced real subset trails FALLNEST in all three metrics, so balancing alone does not explain the gains.The equal-budget comparison also notes that synthetic samples span both classes where random draws may leave ADL regions uncovered.Source: Table 3 · p. 7, Comparison with Equal-Budget Real… · p. 7

Comparison with Full-Data Methods

FALLNEST trades some accuracy for roughly two orders of magnitude fewer training windows, making condensation useful when storing or processing the full corpus is impractical rather than a parity claim.

Comparison with Full-Data Methods

  • Under strict subject-wise splitting, full-data TinyCNN reaches 99.85%, 100.00%, and 99.39%, at the level of published methods, so leakage is not the main explanation for the gap.Published full-data methods report 98–99.8% accuracy, but each published row provides only one metric; BiLSTM–BiGRU reports PR-AUC instead of accuracy.Source: Comparison with Full-Data Methods · p. 7
  • LiteFallNet and BiLSTM–BiGRU use random window-level splits, while and use variants of subject-wise cross-validation.Window-level splits can inflate results through temporal leakage, but the strict-split TinyCNN results show leakage is not the main explanation for the accuracy gap.Source: Comparison with Full-Data Methods · p. 7
  • FALLNEST gives up about 3–6 accuracy points, including 99.3% versus 93.5% on SisFall, while using about two orders of magnitude fewer training windows.Table 5 is indicative only because published protocols differ in splitting, class balance, and metrics.Source: Comparison with Full-Data Methods · p. 7, Table 5 · p. 8
  • FALLNEST does not claim parity with full-data training; condensation targets settings where storing or processing the full corpus is impractical, including rapid prototyping and on-device personalization.The reported accuracy gap is described as mostly the price of 100-fold data reduction.Source: Comparison with Full-Data Methods · pp. 7–8

Ablation Studies

Ablations assess the dual-domain setup, initialization, augmentation, and combined loss; results show dataset-dependent sensitivity, especially on MobiFall, while several design choices remain unisolated.

Ablation Studies

  • On MobiFall, several variants lose 6–20 accuracy points while PR-AUC remains 0.97–0.98; frequency-only instead drops PR-AUC to 0.71.These are single-run ablations (seed 42); validation and test fall prevalence are 14.4% and 25.2%, respectively.Source: Ablation Studies · p. 8, Table 6 · p. 8
  • Frequency-only loses 4.3, 27.3, and 4.7 accuracy points on SisFall, MobiFall, and KFall; time-only loses 3.6, 0.9, and 4.0, respectively.Dual-domain gains are clear on SisFall and KFall, while MobiFall time-only is within seed noise; its frequency-only collapse may reflect 12 spectrogram frames.Source: Ablation Studies · p. 8, Table 6 · p. 8
  • Removing augmentation costs 3.4, 20.0, and 1.6 accuracy points on SisFall, MobiFall, and KFall, respectively.On MobiFall, only 32 samples are condensed, and preserved PR-AUC suggests an operating-point effect.Source: Ablation Studies · p. 8, Table 6 · p. 8
  • Random initialization costs 1.7 and 5.6 accuracy points on SisFall and MobiFall, gains 0.3 on KFall, and lowers SisFall PR-AUC from 0.90 to 0.79.Source: Ablation Studies · p. 8, Table 6 · p. 8
  • Either loss alone is below the combined loss except gradient-only on SisFall within noise; gradient-only beats MMD-only on SisFall and KFall, with the reverse on MobiFall.The ablations do not isolate AVM, normalized versus cosine matching, or clipping, and include no CondTSC or other coreset/condensation baselines; differences from CondTSC are design choices, not validated gains.Source: Ablation Studies · p. 8, Table 6 · p. 8

Training and Inference Cost · Conclusion

FALLNEST condenses imbalanced IMU data into tiny synthetic sets that support strong fall-detection results. The cost analysis distinguishes reduced training data and offline compute from inference costs, and finds condensation pays off only when reusing a condensed set across many model or hyperparameter trials.

Training and Inference Cost

  • Condensation reduces training data and offline compute, not inference latency, memory, or energy, so the paper claims no deployment speedup from it.The paper reports both training and inference costs.Source: Training and Inference Cost · p. 9
  • TinyCNN was best on SisFall and MobiFall and reached 96.6 versus 96.8 for LargeMLP on KFall in the model-scale sweep.The sweep used the same condensed set; LargeMLP differs in architecture as well as size, so it does not isolate parameter count.Source: Training and Inference Cost · p. 9, Table 7 · p. 9
  • MicroCNN was 2.7 points below TinyCNN on SisFall and 30.0 on MobiFall, while SmallCNN and LargeMLP barely exceeded MobiFall’s 74.8% always-ADL floor.On SisFall, the 83.8% floor obscures the gap; Macro-F1 was 0.87 versus 0.91 and is the better indicator.Source: Training and Inference Cost · p. 9, Table 7 · p. 9
  • TinyCNN needs 103–206 M MACs, about 400 KB, and 2.15 ms median CPU latency; the study did not test it on a microcontroller.Latency was measured with single-threaded PyTorch at batch size 1.Source: Training and Inference Cost · p. 9
  • Training on S was 4.6 to 6 times faster than on full T, but condensation plus training took about 90, 24, and 30.5 min versus 37, 18, and 25 min for one full-data run.Condensation pays off only when many architectures or hyperparameters are tried on one condensed set.Source: Training and Inference Cost · p. 9

Conclusion

  • 93.5–95.8% mean accuracy and 0.89–0.98 PR-AUC: a 102 K-parameter TinyCNN trained on 32 to 146 synthetic windows reached these results over five seeds, 3.6 to 6.4 points below full-data training.Evaluation used strict subject-wise splits and naturally imbalanced test sets.Source: Conclusion · p. 9
  • FALLNEST uses class-balanced K-Means initialization, a 7-channel input with AVM, normalized and clipped gradient matching with MMD alignment, and multi-view Fourier augmentation.It distills a large, imbalanced IMU training set into a tiny, class-balanced synthetic set.Source: Conclusion · p. 9
  • K-Means initialization, multi-view augmentation, and the combined loss each mattered in ablations, most clearly when the synthetic set was smallest.Source: Conclusion · p. 9
Loading 2610.18260v1…