TinierHAR: Towards Ultra-Lightweight Deep Learning Models for Efficient Human Activity Recognition on Edge Devices

Daniel Geissler; Mengxi Liu; Paul Lukowicz; Sizhen Bian; Vitor Fortes Rey

REVIEW 4 major objections 5 minor 54 references

TinierHAR: Towards Ultra-Lightweight Deep Learning Models for Efficient Human Activity Recognition on Edge Devices

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read TinierHAR: 43× fewer parameters, 58× fewer MACs, same F1

desk verdict Useful efficiency-first HAR paper; the resource savings are credible, but the F1-parity claim depends on baseline tuning that is not documented. read the letter →

arxiv 2507.07949 v1 pith:3VKGSFE4 submitted 2025-07-10 cs.CV cs.LG

Sizhen Bian , Mengxi Liu , Vitor Fortes Rey , Daniel Geissler , Paul Lukowicz This is my paper · ORCID

classification cs.CVcs.LG

keywords humanactivityrecognitionedgecomputinglightweightdeeplearningdepthwiseseparableconvolutiongatedrecurrentunittemporalaggregationwearablesensorsmodelefficiency

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper sets out to show that a very small deep-learning model can match the accuracy of much larger human activity recognition (HAR) models on wearable data. It claims that its TinierHAR architecture, built from residual depthwise separable convolutions, a bidirectional gated recurrent unit, and attention-based temporal pooling, cuts parameters by 2.7× versus TinyHAR and 43.3× versus DeepConvLSTM, and multiply-accumulate operations (MACs) by 6.4× and 58.6×, while keeping average macro F1-score at parity (1.000× vs TinyHAR, 1.006× vs DeepConvLSTM). If correct, this means memory-constrained edge devices can run competitive HAR without sacrificing recognition quality. The paper also contributes ablation and scaling studies across 14 public datasets that point to which architectural components actually matter.

What carries the argument

The architecture's load-bearing mechanism is the residual depthwise separable convolution block: a depthwise convolution that applies a temporal kernel per input channel followed by a 1×1 pointwise convolution that merges channels, with a residual shortcut and stride-2 max pooling in the first two blocks to halve the temporal dimension. These blocks minimize parameters and MACs while capturing multi-scale spatial patterns. A bidirectional GRU then encodes forward and backward temporal context with weight sharing and no extra cell state, and a learned attention layer weights time steps by importance before classification, replacing static pooling at no extra computational cost. Together these three modules are what the ablation study claims to dissect.

What would settle it

Retrain TinyHAR and DeepConvLSTM on the same 14 datasets with per-model hyperparameter search rather than the shared protocol, and compare macro F1; if either baseline then exceeds TinierHAR by more than about one point on average, the parity claim fails. A complementary check is to deploy all three models on the same edge processor and measure real latency and energy, since MAC and parameter reductions are only practically meaningful if they translate to wall-clock and battery gains.

Watch

Extended reading notes

Core claim

The central claim is that a deliberately minimal combination—residual depthwise separable convolutions for spatial feature extraction, a bidirectional GRU for temporal modeling, and attention-based temporal aggregation—reaches the best accuracy-efficiency trade-off in human activity recognition. Evaluated on 14 public datasets with consistent 4-second windows and leave-one-user-out splits, TinierHAR reduces model parameters by 2.7× relative to TinyHAR and 43.3× relative to DeepConvLSTM, and MACs by 6.4× and 58.6× respectively, while producing average macro F1-scores statistically indistinguishable from both baselines (1.000× and 1.006×). The authors further claim this is the first systematic ablation of spatial-temporal components across TinierHAR, TinyHAR, and DeepConvLSTM, showing that the GRU block is the most critical component in TinierHAR and that convolutional blocks dominate in the other two architectures.

Load-bearing premise

The comparison assumes TinyHAR and DeepConvLSTM were trained to their realistic best with suitable per-model hyperparameters; if they were undertrained, the reported accuracy parity is not a fair accounting.

Editorial extensions

If this is right

On-device HAR with sub-megabyte models becomes feasible: TinierHAR's parameter count makes it deployable in microcontroller-class wearables.
The ablation indicates that GRU layers, not attention or fully connected layers, are the main temporal workhorse; removing self-attention from TinyHAR improved or held F1 on 10 of 14 datasets.
The scaling analysis shows that F1 gains plateau as model width and hidden size grow, so practitioners can choose the smallest configuration that clears a target accuracy.
Depthwise separable convolutions can replace standard convolutions in HAR backbones without accuracy loss, shifting design away from transformer-heavy architectures for tiny devices.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

If the baselines were tuned to their full potential, the parity result would imply that current efficient HAR models still carry significant removable redundancy, an efficiency headroom that TinierHAR exploits.
The finding that removing self-attention helps suggests that for many HAR datasets short-range temporal structure dominates, and attention mechanisms may be unnecessary complexity on edge devices.
A natural test is to combine TinierHAR's convolutional backbone with a frozen pretrained encoder or knowledge distillation to see whether the parameter floor can go lower still.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Quick take: this is a useful, workmanlike extension of the TinyHAR line. The efficiency gains are real and structurally grounded: depthwise separable convolutions, bi-GRU instead of LSTM, and attention pooling. I believe the 2.7x/6.4x and 43.3x/58.6x reductions against TinyHAR and DeepConvLSTM are credible as architecture-dependent measurements. The 14-dataset protocol with five seeds and leave-one-user-out is a solid base, and the open-source release is a plus. The ablation study, especially the component-level analysis across three architectures, is the most original part and will be useful to others designing tiny HAR models.

The soft spots are exactly where the reader flagged them. The abstract claims F1 is 'maintained', but that claim is only as strong as the baseline tuning. Section 4 describes one shared protocol (150 epochs, early stopping, AdamW at 1e-3, five seeds) with no per-model hyperparameter search and no citation of the original DeepConvLSTM implementation. The ablation evidence works against the parity claim: removing the LSTM from DeepConvLSTM rarely hurts, and removing self-attention from TinyHAR often helps. That could be a genuine finding about those components, or it could be a sign that the baselines were run at default settings rather than tuned. Without error bars or significance tests, the 1.000/1.006 ratios are point estimates that could shift with stronger baselines. Also, choosing M=4 and N=16 on the same 14 datasets used for the headline numbers is a mild but real selection-on-test issue.

None of this kills the paper's main value. The efficiency advantage is structural and does not depend on beating the baselines. Even if the baselines could be tuned to match accuracy, the fact that such a small model is within a percentage point is itself informative for deployers. But the accuracy-parity claim, as written, overstates what the evidence supports.

Who should read it: anyone building on-device HAR models who needs a credible lightweight baseline and a map of which architectural components actually earn their keep. It deserves a serious referee at a full venue, provided the authors add error bars, disclose baseline tuning, and cite DeepConvLSTM properly. For this adjunct proceedings, it is acceptable; just do not treat the parity claim as final.

Recommendation: engage with it, cite the efficiency numbers with caveats, and push for the missing statistical details if a revised version crosses your desk.

Referee Report

4 major / 5 minor

Summary. The paper introduces TinierHAR, a HAR architecture combining residual depthwise separable convolutions, a bidirectional GRU, and attention-based temporal aggregation. It evaluates on 14 public HAR datasets, comparing against TinyHAR and DeepConvLSTM. The central claims are that TinierHAR reduces parameters by 2.7x vs. TinyHAR and 43.3x vs. DeepConvLSTM, reduces MACs by 6.4x and 58.6x, respectively, while maintaining average macro F1-scores (1.000x and 1.006x relative to those baselines). The paper also presents a component-wise ablation of all three models and a scaling analysis used to select the final configuration M=4, N=16.

Significance. If the accuracy-parity claim holds, this is a useful contribution: an order-of-magnitude efficiency improvement at matched average F1 on a broad 14-dataset benchmark, with open-source code and a systematic component-level ablation across three architectures. The breadth of the evaluation and the release of all materials are genuine strengths. However, the central parity claim currently rests on baselines whose tuning is not documented and on point estimates without uncertainty quantification, so the significance is conditional on those issues being resolved.

major comments (4)

[Section 4, Figure 2] The shared training protocol (150 epochs, early stopping patience 15, AdamW lr=0.001, scheduler patience 7, five seeds) is applied identically to all models, but no per-model hyperparameter search is reported and the original DeepConvLSTM configuration is not cited. Because the headline claim of 'maintaining the averaged F1-scores' (Figure 2: 1.000x vs. TinyHAR, 1.006x vs. DeepConvLSTM) is a parity claim, it is only meaningful if the baselines are near their realistic best. The manuscript provides no evidence for this; in fact, the DeepConvLSTM ablation (Figure 4) showing that removing the LSTM block often helps is internally consistent with the LSTM being undertrained rather than genuinely unnecessary. Please report per-dataset baseline results from the original papers or a tuned search, and provide standard deviations or confidence intervals for all reported F1 values.
[Section 5.2, Figure 6] The final configuration M=4, N=16 is selected using the same 14 datasets on which the headline efficiency and F1 results are reported. This in-sample model selection makes the reported average F1 and efficiency figures optimistically biased and weakens the claim that this is the 'general optimal choice.' A nested or held-out selection procedure, or at least a clear statement that the headline numbers are affected by selection on the test data, is needed to support an unbiased comparison.
[Section 4, Figure 2] All results are reported as single averages over five seeds, without standard deviations, confidence intervals, or significance tests. The central parity claim (1.000x vs. TinyHAR and 1.006x vs. DeepConvLSTM) and the ablation differences (e.g., -0.95% vs. -1.74% in Figure 3) cannot be distinguished from random variation. Please report per-seed standard deviations or perform paired statistical tests across datasets and seeds.
[Section 4, 'The evaluation metrics include MACs and Parameters'] The method for computing MACs is not specified: no tool (e.g., thop, fvcore), no layer coverage (do the reported MACs include the GRU and attention aggregation?), and no input-shape details are given. Since the efficiency claim is a central contribution, the MACs figures must be reproducible and must include all layers of each model; otherwise the 58.6x reduction vs. DeepConvLSTM could be inflated by an incomplete or inconsistent counting method.

minor comments (5)

[Section 5, first paragraph] The word 'desribed' should be 'described'.
[Section 5.2, first paragraph] The text lists 'DG, USCHAD, RM, and RECGYM' but Figure 6 appears to show RW, not RM; this looks like a typo and should be corrected.
[Figure 2 caption] The caption uses 'Tinierhar' with inconsistent capitalization; the model name should be 'TinierHAR' throughout.
[Page headers] The page header still reads 'Conference’17, July 2017, Washington, DC, USA', which is a leftover template from a previous venue and should be updated to the actual publication venue.
[Section 5.1] The ablation descriptions say a block was 'removed' but do not specify what replaced it, e.g., when the GRU block is removed, does the temporal aggregation receive the CNN output directly, and when the aggregation block is removed, is average pooling used? Clarifying these replacements is needed for reproducibility.

Circularity Check

0 steps flagged · score 1.0 of 10

No circularity found: efficiency numbers are structural, F1 parity is an external benchmark, and the flagged baseline-tuning and M/N-selection issues are validity concerns, not fitted-input-called-prediction.

full rationale

The paper's two central claims are (1) reduced parameters and MACs versus TinyHAR and DeepConvLSTM, and (2) maintained average F1. Both are supported by direct measurements, not by a derivation that reduces to its own inputs. The parameter/MAC reductions follow from the architecture definitions and are compared against external, well-known baseline architectures; there is no equation in which the claimed outcome is inserted by construction. The F1 comparison is an empirical benchmark over 14 public datasets with a shared training protocol, not a fitted parameter renamed as a prediction. The ablation studies are empirical decompositions of the architectures, and no 'uniqueness theorem' or load-bearing self-citation is invoked to force the design. The paper does cite prior work by overlapping DFKI authors, but those citations appear in context-setting and general HAR references; the core comparison rests on public datasets and external baselines, not on the authors' own prior results. The reviewer-flagged weaknesses--no per-model hyperparameter search for baselines and the selection of M=4, N=16 based on 'the global optimized scale across the 14 datasets' before reporting results on those same datasets--are legitimate threats to the fairness and generalizability of the benchmark, but they are not circularity: the reported F1 scores are measurements, not predictions derived from the fitted hyperparameters. No passage admits to a circular step, and none was found. Therefore the circularity score is low, reflecting only the minor presence of author self-citations in non-load-bearing roles.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

The headline results rest on two hand-chosen hyperparameters (M, N), a single windowing convention, and the faithfulness of baseline re-implementations. No new physical entities or forces are introduced.

free parameters (3)

Residual separable conv blocks (M) = 4
Selected as the global optimal configuration across the 14 evaluation datasets (Section 5.2); the reported efficiency and accuracy numbers are for this size.
GRU hidden size (N) = 16
Selected together with M in Section 5.2; no held-out validation is reported for this choice.
Sliding window length = 4 seconds with 2 second overlap
Applied uniformly to all 14 datasets (Table 1); window length affects input dimensions and therefore parameter counts and MACs.

assumptions (3)

domain assumption The 14 public datasets and the leave-one-subject-out splits give a representative and unbiased estimate of HAR performance.
Section 4, Experimental Results; the paper generalizes across these datasets without external validation.
domain assumption TinyHAR and DeepConvLSTM baselines were implemented faithfully and trained under comparable conditions.
Section 4 gives a shared training protocol but no per-baseline tuning details and no citation for DeepConvLSTM, so faithful implementation is assumed.
domain assumption Depthwise separable convolutions preserve the representational capacity needed for HAR feature extraction.
Section 3 replaces standard convolutions with depthwise separable blocks; the accuracy parity claim depends on this capacity assumption.

how reviews work

0 comments

Cite this review

Pith. "Pith review of TinierHAR: Towards Ultra-Lightweight Deep Learning Models for Efficient Human Activity Recognition on Edge Devices." pith.science (2026). https://pith.science/paper/3VKGSFE4

@misc{pith2026250707949,
  author       = {Pith},
  title        = {Pith review of: TinierHAR: Towards Ultra-Lightweight Deep Learning Models for Efficient Human Activity Recognition on Edge Devices},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3VKGSFE4}},
  note         = {Machine review of arXiv:2507.07949}
}

read the original abstract

Human Activity Recognition (HAR) on resource-constrained wearable devices demands inference models that harmonize accuracy with computational efficiency. This paper introduces TinierHAR, an ultra-lightweight deep learning architecture that synergizes residual depthwise separable convolutions, gated recurrent units (GRUs), and temporal aggregation to achieve SOTA efficiency without compromising performance. Evaluated across 14 public HAR datasets, TinierHAR reduces Parameters by 2.7x (vs. TinyHAR) and 43.3x (vs. DeepConvLSTM), and MACs by 6.4x and 58.6x, respectively, while maintaining the averaged F1-scores. Beyond quantitative gains, this work provides the first systematic ablation study dissecting the contributions of spatial-temporal components across proposed TinierHAR, prior SOTA TinyHAR, and the classical DeepConvLSTM, offering actionable insights for designing efficient HAR systems. We finally discussed the findings and suggested principled design guidelines for future efficient HAR. To catalyze edge-HAR research, we open-source all materials in this work for future benchmarking\footnote{https://github.com/zhaxidele/TinierHAR}

Figures

Figures reproduced from arXiv: 2507.07949 by the authors.

**Figure 2.** Evaluation result of TinierHAR on 14 HAR datasets. Averages across datasets: 2.7x, 43.3x in parameters, 6.4x, 58.6x in MACs, 1.000x, [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** F1-Score Variation of TinierHAR Ablation Study Across Datasets. Averaged variations are -9.19%, -0.95%, -19.63%, -1.74% when removing [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Figures from the paper (3 more)

**Figure 4.** Figure 4: F1-Score Variation of DeepConvLSTM Ablation Study Across Datasets. Averaged variations are -27.50%, -1.84% when removing the [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]

**Figure 5.** Figure 5: F1-Score Variation of TinyHAR Ablation Study Across Datasets. Averaged variations are -55.06%, +0.25%, +0.40%, -5.26%, -0.02% when [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]

**Figure 6.** Figure 6: Metrics Across Models Configurations and Datasets (M_N of the labels in bottom x-axis means M residual separable convolutional [PITH_FULL_IMAGE:figures/full_fig_p005_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

54 extracted references · 45 canonical work pages

[1]

https://archive.ics.uci.edu/dataset/245/daphnet+freezing+of+gait

DG Dataset. https://archive.ics.uci.edu/dataset/245/daphnet+freezing+of+gait. Accessed: 2025-04-30

work page 2025
[2]

https://archive.ics.uci.edu/dataset/256/daily+and+sports+ activities

DSADS Dataset. https://archive.ics.uci.edu/dataset/256/daily+and+sports+ activities. Accessed: 2025-04-30

work page 2025
[3]

https://archive.ics.uci.edu/dataset/341/smartphone+based+ recognition+of+human+activities+and+postural+transitions

HAPT Dataset. https://archive.ics.uci.edu/dataset/341/smartphone+based+ recognition+of+human+activities+and+postural+transitions. Accessed: 2025-04- 30

work page 2025
[4]

https://archive.ics.uci.edu/dataset/319/mhealth+dataset

MHEALTH Dataset. https://archive.ics.uci.edu/dataset/319/mhealth+dataset. Accessed: 2025-04-30

work page 2025
[5]

https://www.kaggle.com/datasets/malekzadeh/ motionsense-dataset

MOTIONSENSE Dataset. https://www.kaggle.com/datasets/malekzadeh/ motionsense-dataset. Accessed: 2025-04-30

work page 2025
[6]

https://archive.ics.uci.edu/dataset/226/opportunity+activity+ recognition

OPPO Dataset. https://archive.ics.uci.edu/dataset/226/opportunity+activity+ recognition. Accessed: 2025-04-30

work page 2025
[7]

https://archive.ics.uci.edu/dataset/231/pamap2+physical+ activity+monitoring

PAMAP2 Dataset. https://archive.ics.uci.edu/dataset/231/pamap2+physical+ activity+monitoring. Accessed: 2025-04-30

work page 2025
[8]

https://zhaxidele.github.io/RecGym/

RECGYM Dataset. https://zhaxidele.github.io/RecGym/. Accessed: 2025-04-30

work page 2025

Show all 54 references

[9]

https://www.uni-mannheim.de/dws/research/projects/activity- recognition/dataset/dataset-realworld/

RW Dataset. https://www.uni-mannheim.de/dws/research/projects/activity- recognition/dataset/dataset-realworld/. Accessed: 2025-04-30

2025
[10]

https://www.utwente.nl/en/eemcs/ps/research/dataset/

SHO Dataset. https://www.utwente.nl/en/eemcs/ps/research/dataset/. Accessed: 2025-04-30

2025
[11]

http://har-dataset.org/doku.php?id=wiki:dataset

SKODAR Dataset. http://har-dataset.org/doku.php?id=wiki:dataset. Accessed: 2025-04-30

2025
[12]

https://archive.ics.uci.edu/dataset/240/human+activity+ recognition+using+smartphones

UCI Dataset. https://archive.ics.uci.edu/dataset/240/human+activity+ recognition+using+smartphones. Accessed: 2025-04-30

2025
[13]

https://sipi.usc.edu/had/

USCHAD Dataset. https://sipi.usc.edu/had/. Accessed: 2025-04-30

2025
[14]

https://archive.ics.uci.edu/dataset/507/wisdm+smartphone+ and+smartwatch+activity+and+biometrics+dataset

WISDM Dataset. https://archive.ics.uci.edu/dataset/507/wisdm+smartphone+ and+smartwatch+activity+and+biometrics+dataset. Accessed: 2025-04-30

2025
[15]

Oliver Amft, Luis Ignacio Lopera González, Paul Lukowicz, Sizhen Bian, and Paul Burggraf. 2020. Wearables to fight COVID-19: From symptom tracking to contact tracing. IEEE Pervasive Computing 19, 4 (2020), 53–60

2020
[16]

Sheeza Batool, Muhammad Hassan Khan, and Muhammad Shahid Farid. 2024. An ensemble deep learning model for human activity analysis using wearable sensory data. Applied Soft Computing 159 (2024), 111599

2024
[17]

Sizhen Bian, Pixi Kang, Julian Moosmann, Mengxi Liu, Pietro Bonazzi, Roman Rosipal, and Michele Magno. 2024. On-device Learning of EEGNet-based Network For Wearable Motor Imagery Brain-Computer Interface. In Proceedings of the 2024 ACM International Symposium on Wearable Compu...

2024
[18]

Sizhen Bian and Paul Lukowicz. 2021. Capacitive sensing based on-board hand gesture recognition with tinyml. In Adjunct Proceedings of the 2021 ACM Interna- tional Joint Conference on Pervasive and Ubiquitous Computing and Proceedings of the 2021 ACM International Symposium on...

2021
[19]

Sizhen Bian and Paul Lukowicz. 2021. A systematic study of the influence of various user specific and environmental factors on wearable human body capacitance sensing. In EAI International Conference on Body Area Networks . Springer, 247–274

2021
[20]

Sizhen Bian, Vitor F Rey, Peter Hevesi, and Paul Lukowicz. 2019. Passive capacitive based approach for full body gym workout recognition and counting. In2019 IEEE International Conference on Pervasive Computing and Communications (PerCom . IEEE, 1–10

2019
[21]

Sizhen Bian, Vitor Fortes Rey, Siyu Yuan, and Paul Lukowicz. 2025. Hybrid CNN-Dilated Self-attention Model Using Inertial and Body-Area Electrostatic Sensing for Gym Workout Recognition, Counting, and User Authentification. arXiv preprint arXiv:2503.06311 (2025)

2025 arXiv
[22]

Pietro Bonazzi, Sizhen Bian, Giovanni Lippolis, Yawei Li, Sadique Sheik, and Michele Magno. 2024. Retina: Low-power eye tracking with event camera and spiking hardware. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 5684–5692

2024
[23]

Shizhuo Deng, Jiaqi Chen, Da Teng, Chuangui Yang, Dongyue Chen, Tong Jia, and Hao Wang. 2023. Lhar: Lightweight human activity recognition on knowledge distillation. IEEE Journal of Biomedical and Health Informatics (2023)

2023
[24]

Stéphane d’Ascoli, Hugo Touvron, Matthew L Leavitt, Ari S Morcos, Giulio Biroli, and Levent Sagun. 2021. Convit: Improving vision transformers with soft convolutional inductive biases. In International conference on machine learning . PMLR, 2286–2296

2021
[25]

Laura Fiorini, Federica G Cornacchia Loizzo, Alessandra Sorrentino, Jaeseok Kim, Erika Rovini, Alessandro Di Nuovo, and Filippo Cavallo. 2021. Daily gesture recognition during human-robot interaction combining vision and wearable systems. IEEE Sensors Journal 21, 20 (2021), 23...

2021
[26]

Daniel Geissler, Dominique Nshimyimana, Vitor Fortes Rey, Sungho Suh, Bo Zhou, and Paul Lukowicz. 2024. Beyond Confusion: A Fine-grained Dialectical Examination of Human Activity Recognition Benchmark Datasets. arXiv preprint arXiv:2412.09037 (2024)

2024 arXiv
[27]

Rebeen Ali Hamad, Masashi Kimura, Longzhi Yang, Wai Lok Woo, and Bo Wei
[28]

Tom Hoddes, Alex Bijamov, Saket Joshi, Daniel Roggen, Ali Etemad, Robert Harle, and David Racz. 2025. Scaling laws in wearable human activity recognition.arXiv preprint arXiv:2502.03364 (2025)

2025 arXiv
[29]

Zhiqing Hong, Zelong Li, Shuxin Zhong, Wenjun Lyu, Haotian Wang, Yi Ding, Tian He, and Desheng Zhang. 2024. Crosshar: Generalizing cross-dataset human activity recognition via hierarchical self-supervised pretraining. Proceedings of the ACM on Interactive, Mobile, Wearable and...

2024
[30]

Ogbonna Michael Igwe, Yi Wang, George C Giakos, and Jian Fu. 2022. Human activity recognition in smart environments employing margin setting algorithm. Journal of Ambient Intelligence and Humanized Computing 13, 7 (2022), 3669– 3681

2022
[31]

Hari Kang, Donghyun Kim, and Kar-Ann Toh. 2025. Human Activity Recognition Through Augmented WiFi CSI Signals by Lightweight Attention-GRU. Sensors 25, 5 (2025), 1547

2025
[32]

Pixi Kang, Julian Moosmann, Sizhen Bian, and Michele Magno. 2024. On-Device Training Empowered Transfer Learning For Human Activity Recognition. arXiv preprint arXiv:2407.03644 (2024)

2024 arXiv
[33]

Pooja Lalwani and R Ganeshan. 2024. A Novel CNN-BiLSTM-GRU Hybrid Deep Learning Model for Human Activity Recognition. International Journal of Computational Intelligence Systems 17, 1 (2024), 1–20

2024
[34]

Emanuele Lattanzi, Lorenzo Calisti, and Chiara Contoli. 2024. Are Transformers a Useful Tool for Tiny devices in Human Activity Recognition?. InProceedings of the 2024 8th International Conference on Advances in Artificial Intelligence . 339–344

2024
[35]

Won-Seon Lim, Wangduk Seo, Dae-Won Kim, and Jaesung Lee. 2023. Efficient Human Activity Recognition Using Lookup Table-Based Neural Architecture Search for Mobile Devices. IEEE Access 11 (2023), 71727–71738

2023
[36]

Mengxi Liu, Sizhen Bian, and Paul Lukowicz. 2022. Non-contact, real-time eye blink detection with capacitive sensing. In Proceedings of the 2022 ACM International Symposium on Wearable Computers . 49–53

2022
[37]

Mengxi Liu, Daniel Geißler, Sizhen Bian, Bo Zhou, and Paul Lukowicz. 2025. Assessing the Impact of Sampling Irregularity in Time Series Data: Human Activity Recognition As A Case Study. arXiv preprint arXiv:2501.15330 (2025)

2025 arXiv
[38]

Victor Luder, Sizhen Bian, and Michele Magno. 2024. CycloWatt: An Afford- able, TinyML-enhanced IoT Device Revolutionizing Cycling Power Metrics. In 2024 IEEE International Instrumentation and Measurement Technology Conference (I2MTC). IEEE, 1–6

2024
[39]

Ala Mhalla and Jean-Marie Favreau. 2024. Domain adaptation framework for per- sonalized human activity recognition models. Multimedia Tools and Applications 83, 25 (2024), 66775–66797

2024
[40]

Shenghuan Miao, Ling Chen, and Rong Hu. 2024. Spatial-temporal masked autoencoder for multi-device wearable human activity recognition. Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies7, 4 (2024), 1–25

2024
[41]

Aimé Cedric Muhoza, Emmanuel Bergeret, Corinne Brdys, and Francis Gary. 2023. Power consumption reduction for IoT devices thanks to Edge-AI: Application to human activity recognition. Internet of Things 24 (2023), 100930

2023
[42]

Nabasmita Phukan, Shailesh Mohine, Achinta Mondal, M Sabarimalai Manikan- dan, and Ram Bilas Pachori. 2022. Convolutional neural network-based human activity recognition for edge fitness and context-aware health monitoring devices. IEEE Sensors Journal 22, 22 (2022), 21816–21826

2022
[43]

Mutegeki Ronald, Alwin Poulose, and Dong Seog Han. 2021. iSPLInception: an inception-ResNet deep learning architecture for human activity recognition.IEEE Access 9 (2021), 68985–69001

2021
[44]

Fatemeh Serpush, Mohammad Bagher Menhaj, Behrooz Masoumi, and Babak Karasfi. 2022. Wearable sensor-based human activity recognition in the smart healthcare system. Computational intelligence and neuroscience 2022, 1 (2022), 1391906

2022
[45]

Sibo Song, Ngai-Man Cheung, Vijay Chandrasekhar, and Bappaditya Mandal
[46]

Megha Thukral, Harish Haresamudram, and Thomas Ploetz. 2025. Cross-Domain HAR: Few-Shot Transfer Learning for Human Activity Recognition. ACM Trans- actions on Intelligent Systems and Technology 16, 1 (2025), 1–35

2025
[47]

Yang Xu and Ting Ting Qiu. 2021. Human activity recognition and embedded ap- plication based on convolutional neural network. Journal of Artificial Intelligence and Technology 1, 1 (2021), 51–60

2021
[48]

Myung-Kyu Yi, Wai-Kong Lee, and Seong Oun Hwang. 2023. A human activity recognition method based on lightweight feature extraction combined with pruned and quantized CNN for wearable device. IEEE Transactions on Consumer Electronics 69, 3 (2023), 657–670

2023
[49]

Shibo Zhang, Yaxuan Li, Shen Zhang, Farzad Shahabi, Stephen Xia, Yu Deng, and Nabil Alshurafa. 2022. Deep learning in human activity recognition with wearable sensors: A review on advances. Sensors 22, 4 (2022), 1476

2022
[50]

Yexu Zhou, Tobias King, Haibin Zhao, Yiran Huang, Till Riedel, and Michael Beigl. 2024. MLP-HAR: Boosting Performance and Efficiency of HAR Models on Edge Devices with Purely Fully Connected Layers. In Proceedings of the 2024 ACM International Symposium on Wearable Computers ....

2024
[51]

Yexu Zhou, Haibin Zhao, Yiran Huang, Till Riedel, Michael Hefenbrock, and Michael Beigl. 2022. Tinyhar: A lightweight deep learning model designed for human activity recognition. In Proceedings of the 2022 ACM International Symposium on Wearable Computers. 89–93

2022
[52]

Jianping Zhu, Xin Lou, and Wenbin Ye. 2021. Lightweight deep learning model in mobile-edge computing for radar-based human activity recognition. IEEE Internet of Things Journal 8, 15 (2021), 12350–12359

2021
[2018]

In Proceedings of the 26th ACM international conference on Multimedia

Deep adaptive temporal pooling for activity recognition. In Proceedings of the 26th ACM international conference on Multimedia . 1829–1837
[2021]

Neural Computing and Applications 33 (2021), 13705–13722

Dilated causal convolution with multi-head self attention for sensor human activity recognition. Neural Computing and Applications 33 (2021), 13705–13722

2021

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

https://archive.ics.uci.edu/dataset/245/daphnet+freezing+of+gait

DG Dataset. https://archive.ics.uci.edu/dataset/245/daphnet+freezing+of+gait. Accessed: 2025-04-30

work page 2025

[2] [2]

https://archive.ics.uci.edu/dataset/256/daily+and+sports+ activities

DSADS Dataset. https://archive.ics.uci.edu/dataset/256/daily+and+sports+ activities. Accessed: 2025-04-30

work page 2025

[3] [3]

https://archive.ics.uci.edu/dataset/341/smartphone+based+ recognition+of+human+activities+and+postural+transitions

HAPT Dataset. https://archive.ics.uci.edu/dataset/341/smartphone+based+ recognition+of+human+activities+and+postural+transitions. Accessed: 2025-04- 30

work page 2025

[4] [4]

https://archive.ics.uci.edu/dataset/319/mhealth+dataset

MHEALTH Dataset. https://archive.ics.uci.edu/dataset/319/mhealth+dataset. Accessed: 2025-04-30

work page 2025

[5] [5]

https://www.kaggle.com/datasets/malekzadeh/ motionsense-dataset

MOTIONSENSE Dataset. https://www.kaggle.com/datasets/malekzadeh/ motionsense-dataset. Accessed: 2025-04-30

work page 2025

[6] [6]

https://archive.ics.uci.edu/dataset/226/opportunity+activity+ recognition

OPPO Dataset. https://archive.ics.uci.edu/dataset/226/opportunity+activity+ recognition. Accessed: 2025-04-30

work page 2025

[7] [7]

https://archive.ics.uci.edu/dataset/231/pamap2+physical+ activity+monitoring

PAMAP2 Dataset. https://archive.ics.uci.edu/dataset/231/pamap2+physical+ activity+monitoring. Accessed: 2025-04-30

work page 2025

[8] [8]

https://zhaxidele.github.io/RecGym/

RECGYM Dataset. https://zhaxidele.github.io/RecGym/. Accessed: 2025-04-30

work page 2025