pith. sign in

arxiv: 2510.24738 · v2 · submitted 2025-10-14 · 📡 eess.SP · cs.LG

StrikeWatch: Wrist-worn Gait Recognition with Compact Time-series Models on Low-power FPGAs

Tianheng Ling , Chao Qian , Peter Zdankin , Torben Weis , Gregor Schiele This is my paper

Pith reviewed 2026-05-18 06:59 UTC · model grok-4.3

classification 📡 eess.SP cs.LG
keywords gait recognitionwrist IMUFPGA deploymentdeep learningreal-time inferenceenergy efficient wearablesheel strike detectionon-device processing
0
0 comments X

The pith

A 6-bit quantized 1D separable CNN on a low-power FPGA wrist device classifies heel versus forefoot strikes from IMU data at 0.847 average F1 score.

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

This paper shows that real-time detection of running gait patterns can happen entirely on a small wrist-worn device by processing IMU signals locally. Previous systems needed cameras, insoles, or cloud links and could not give feedback while running. The authors test four compact neural network designs and map them to two low-power FPGAs, finding that a quantized separable CNN delivers the best mix of accuracy and speed. Outdoor tests with twelve runners confirm the model runs fast enough and uses so little energy that a small battery lasts many days. The result opens a path to immediate visual or sound cues that help runners fix harmful foot-strike habits without extra hardware.

Core claim

The paper claims that a 6-bit quantized 1D separable convolutional neural network running on the Lattice iCE40UP5K FPGA can classify wrist IMU signals as heel or forefoot strikes with an average F1 score of 0.847. The same configuration uses 0.350 microjoules per inference and finishes each prediction in 0.140 milliseconds when clocked at 20 MHz, allowing continuous operation for 13.6 days on a 320 mAh battery. The evaluation uses a custom hardware prototype and a labeled dataset gathered during outdoor runs by twelve participants.

What carries the argument

The 1D-SepCNN (one-dimensional separable convolutional neural network) quantized to 6 bits, which extracts features from sequential IMU time-series data and maps them to compact FPGA logic for low-energy classification of gait strike type.

If this is right

  • Runners obtain immediate visual or auditory feedback on foot strike during outdoor runs to reduce injury risk.
  • The measured energy use supports weeks of continuous inference on a typical small wearable battery.
  • Hardware-aware quantization and model compression enable real-time performance on resource-limited FPGAs.
  • Clear accuracy-versus-power trade-offs appear between the four tested architectures and the two FPGA platforms.
  • The same pipeline can support other on-device time-series tasks such as activity detection in wearables.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Commercial smartwatches could adopt similar models to add gait feedback without extra sensors or cloud calls.
  • Retraining or domain adaptation may be needed for indoor running or large speed changes not covered in the outdoor dataset.
  • Longer user studies could test whether real-time cues produce lasting changes in running form.
  • Combining the IMU model with heart-rate or GPS data might raise accuracy while staying inside the same power budget.

Load-bearing premise

Data from twelve outdoor runners on the custom prototype will let the models keep high accuracy for new users, different surfaces, and varying speeds.

What would settle it

New tests with runners on varied surfaces and speeds that show the F1 score falling below 0.7 would show the models do not generalize as claimed.

Figures

Figures reproduced from arXiv: 2510.24738 by Chao Qian, Gregor Schiele, Peter Zdankin, Tianheng Ling, Torben Weis.

Figure 1
Figure 1. Figure 1: Wrist-worn StrikeWatch system for real-time heel strike recognition, providing on-device visual (LED) and auditory feedback. Runner silhouettes adapted from [11]. To address these limitations while enabling autonomous edge intelligence for IoT, this study proposes StrikeWatch, a wrist-worn system that integrates software-hardware co-design for on-device gait recognition in practical outdoor settings (see … view at source ↗
Figure 2
Figure 2. Figure 2: StrikeWatch hardware and schematic diagram As shown in [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 4
Figure 4. Figure 4: Sample triaxial wrist-worn acceleration signals for [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Validation F1 score versus energy consumption for deployable configurations of four models on the XC7S15 FPGA. [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Cross-participant distribution of quantized test F1 Scores for each selected Model configuration from Table II [PITH_FULL_IMAGE:figures/full_fig_p009_6.png] view at source ↗
read the original abstract

Running offers substantial health benefits, but improper gait patterns can lead to injuries, particularly without expert feedback. While prior gait analysis systems based on cameras, insoles, or body-mounted sensors have demonstrated effectiveness, they are often bulky and limited to offline, post-run analysis. Wrist-worn wearables offer a more practical and non-intrusive alternative, yet enabling real-time gait recognition on such devices remains challenging due to noisy Inertial Measurement Unit (IMU) signals, limited computing resources, and dependence on cloud connectivity. This paper introduces StrikeWatch, a compact wrist-worn system that performs entirely on-device, real-time gait recognition using IMU signals. As a case study, we target the detection of heel versus forefoot strikes to enable runners to self-correct harmful gait patterns through visual and auditory feedback during running. We propose four compact DL architectures (1D-CNN, 1D-SepCNN, LSTM, and Transformer) and optimize them for energy-efficient inference on two representative embedded Field-Programmable Gate Arrays (FPGAs): the AMD Spartan-7 XC7S15 and the Lattice iCE40UP5K. Using our custom-built hardware prototype, we collect a labeled dataset from outdoor running sessions and evaluate all models via a fully automated deployment pipeline. Our results reveal clear trade-offs between model complexity and hardware efficiency. Evaluated across 12 participants, 6-bit quantized 1D-SepCNN achieves the highest average F1 score of 0.847 while consuming just 0.350 microjoule per inference with a latency of 0.140 ms on the iCE40UP5K running at 20 MHz. This configuration supports up to 13.6 days of continuous inference on a 320 mAh battery. All datasets and code are available in the GitHub repository https://github.com/tianheng-ling/StrikeWatch.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 2 minor

Summary. The paper presents StrikeWatch, a wrist-worn IMU-based system for real-time on-device detection of heel versus forefoot strikes to provide gait feedback during running. It introduces four compact time-series models (1D-CNN, 1D-SepCNN, LSTM, Transformer), quantizes them to low bit-widths, and deploys them via an automated pipeline on two low-power FPGAs (AMD Spartan-7 XC7S15 and Lattice iCE40UP5K). A custom hardware prototype collects a labeled dataset from 12 outdoor runners; the headline result is that 6-bit quantized 1D-SepCNN achieves 0.847 average F1 while using 0.350 μJ per inference and 0.140 ms latency on the iCE40UP5K at 20 MHz, supporting multi-day battery life. Datasets and code are released.

Significance. If the reported performance generalizes under proper subject-independent validation, the work provides a concrete demonstration of energy-efficient, fully on-device gait analysis on resource-constrained FPGAs, advancing practical wearable systems for injury prevention. The open release of the dataset, code, and automated deployment pipeline is a clear strength that enables reproducibility and follow-on research.

major comments (1)
  1. [Experiments / Results section] Experiments / Results section: The manuscript reports an average F1 of 0.847 for the 6-bit 1D-SepCNN 'evaluated across 12 participants' but does not specify the train/test partitioning protocol (e.g., leave-one-subject-out, per-user folds, or pooled mixed-subject splits). Because the central deployment claim requires reliable performance on unseen runners without per-user calibration, the absence of this detail makes it impossible to determine whether the headline metric reflects inter-user generalization or benefits from subject overlap.
minor comments (2)
  1. [Abstract] The abstract states that 'all datasets and code are available' but the main text does not repeat the exact persistent repository link or commit hash.
  2. [Figures] Figure captions for the hardware prototype and FPGA resource utilization could include explicit power and latency annotations to improve immediate readability.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for their careful review and constructive feedback. We address the single major comment below and will update the manuscript to ensure the evaluation protocol is fully transparent.

read point-by-point responses

  1. Referee: The manuscript reports an average F1 of 0.847 for the 6-bit 1D-SepCNN 'evaluated across 12 participants' but does not specify the train/test partitioning protocol (e.g., leave-one-subject-out, per-user folds, or pooled mixed-subject splits). Because the central deployment claim requires reliable performance on unseen runners without per-user calibration, the absence of this detail makes it impossible to determine whether the headline metric reflects inter-user generalization or benefits from subject overlap.

    Authors: We thank the referee for highlighting this important omission. The reported results were obtained using leave-one-subject-out cross-validation: for each of the 12 participants the model was trained exclusively on data from the remaining 11 participants and evaluated on the held-out participant; the final 0.847 F1 is the average across the 12 folds. This protocol guarantees zero subject overlap between training and test sets and directly supports the claim of reliable performance on unseen runners without per-user calibration. We will add an explicit description of the LOSO partitioning, including the exact fold structure, to the Experiments / Results section in the revised manuscript. revision: yes

Circularity Check

0 steps flagged

No circularity: results are direct empirical measurements

full rationale

The paper reports model F1 scores, energy, and latency from training compact DL architectures on a custom outdoor IMU dataset collected from 12 participants, followed by direct hardware measurements on two FPGAs. No derivation chain, first-principles prediction, or fitted parameter is presented that reduces by construction to its own inputs; all central claims are post-training empirical evaluations on held-out data rather than self-referential equations or self-citation load-bearing steps. The evaluation protocol details (e.g., subject splits) affect generalization strength but do not create circularity under the defined criteria.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central claim rests on standard supervised learning assumptions plus the domain premise that wrist IMU contains usable strike-type information; no new physical entities are postulated and the only notable free parameter is the 6-bit quantization width chosen for hardware fit.

free parameters (1)
  • Quantization bit-width
    6-bit width selected to trade accuracy against FPGA resource usage and energy; value is not derived from first principles.
axioms (1)
  • domain assumption Wrist-mounted IMU signals contain sufficient discriminative information for heel versus forefoot strike classification
    Invoked when the models are trained and evaluated on the collected dataset; treated as validated by the reported F1 scores.

pith-pipeline@v0.9.0 · 5890 in / 1253 out tokens · 38441 ms · 2026-05-18T06:59:27.435338+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

26 extracted references · 26 canonical work pages · 1 internal anchor

  1. [1]

    Is this the real life, or is this just laboratory? A scoping review of IMU-based running gait analysis,

    L. C. Benson, A. M. R ¨ais¨anen, C. A. Clermont, and R. Ferber, “Is this the real life, or is this just laboratory? A scoping review of IMU-based running gait analysis,”Sensors, vol. 22, no. 5, p. 1722, 2022

    work page 2022

  2. [2]

    Risk factors for injuries in runners: A systematic review of foot strike technique and its classification at impact,

    A. Burke, S. Dillon, S. O’Connor, E. F. Whyte, S. Gore, and K. A. Moran, “Risk factors for injuries in runners: A systematic review of foot strike technique and its classification at impact,”Orthopaedic journal of sports medicine, vol. 9, no. 9, 2021

    work page 2021

  3. [3]

    F-VESPA: A kinematic-based algo- rithm for real-time heel-strike detection during walking,

    C. Karakasis and P. Artemiadis, “F-VESPA: A kinematic-based algo- rithm for real-time heel-strike detection during walking,” inInternational Conference on Intelligent Robots and Systems. IEEE, 2021

    work page 2021

  4. [4]

    A study on real-time visualizations during sports activities on smartwatches,

    A. Schiewe, A. Krekhov, F. Kerber, F. Daiber, and J. Kr ¨uger, “A study on real-time visualizations during sports activities on smartwatches,” inProceedings of the 19th International Conference on Mobile and Ubiquitous Multimedia, 2020, pp. 18–31

    work page 2020

  5. [5]

    Footstriker: An EMS-based foot strike assistant for running,

    M. Hassan, F. Daiber, F. Wiehr, F. Kosmalla, and A. Kr¨uger, “Footstriker: An EMS-based foot strike assistant for running,”Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies, vol. 1, no. 1, pp. 1–18, 2017

    work page 2017

  6. [6]

    Design and implementation of wearable IMU sensor system for heel-strike and toe-off gait parameter measurement,

    M. Muhamad, A. Razak, A. Halim, M. M. Idros, F. Osman, S. Al Junid, and S. P. Chee, “Design and implementation of wearable IMU sensor system for heel-strike and toe-off gait parameter measurement,” inIn- ternational Conference on Applied Electronics and Engineering. IEEE, 2023, pp. 1–5

    work page 2023

  7. [7]

    Bespoke fuzzy logic design to automate a better understanding of running gait analysis,

    F. Young, S. Stuart, R. McNicol, R. Morris, C. Downs, M. Coleman, and A. Godfrey, “Bespoke fuzzy logic design to automate a better understanding of running gait analysis,”Journal of Biomedical and Health Informatics, vol. 27, no. 5, pp. 2178–2185, 2022

    work page 2022

  8. [8]

    Machine Learning-based determination of gait events from foot-mounted inertial units,

    M. Zago, M. Tarabini, M. Delfino Spiga, C. Ferrario, F. Bertozzi, C. Sforza, and M. Galli, “Machine Learning-based determination of gait events from foot-mounted inertial units,”Sensors, vol. 21, p. 839, 2021

    work page 2021

  9. [9]

    Human activity recognition in smart cities from smart watch data using LSTM Recurrent Neural Networks,

    M. Kandpal, B. Sharma, R. K. Barik, S. Chowdhury, S. S. Patra, and I. B. Dhaou, “Human activity recognition in smart cities from smart watch data using LSTM Recurrent Neural Networks,” in1st International Conference on Advanced Innovations in Smart Cities. IEEE, 2023

    work page 2023

  10. [10]

    Personalized gait detection using a wrist-worn accelerometer,

    G. Cola, M. Avvenuti, F. Musso, and A. Vecchio, “Personalized gait detection using a wrist-worn accelerometer,” inInternational Conference on Wearable and Implantable Body Sensor Networks. IEEE, 2017

    work page 2017

  11. [11]

    Foot strike position & their effects in running: Part 2,

    HealthyStep, “Foot strike position & their effects in running: Part 2,” ht tps://www.healthystep.co.uk/advice/foot-strike-position-affects-part-2/, 2025, accessed: 2025-05-29

    work page 2025

  12. [12]

    Unsupervised segmentation of heel-strike IMU data using rapid cluster estimation of wavelet features,

    M. Yuwono, S. W. Su, B. D. Moulton, and H. T. Nguyen, “Unsupervised segmentation of heel-strike IMU data using rapid cluster estimation of wavelet features,” in35th Annual International Conference of the IEEE Engineering in Medicine and Biology Society. IEEE, 2013

    work page 2013

  13. [13]

    Identification of footstrike pattern using accelerometry and Machine Learning,

    J. M. Mahoney, M. B. Rhudy, J. Outerleys, I. S. Davis, and A. R. Altman-Singles, “Identification of footstrike pattern using accelerometry and Machine Learning,”Journal of Biomechanics, vol. 174, 2024

    work page 2024

  14. [14]

    Esti- mation of fine-grained foot strike patterns with wearable smartwatch devices,

    H. Joo, H. Kim, J.-K. Ryu, S. Ryu, K.-M. Lee, and S.-C. Kim, “Esti- mation of fine-grained foot strike patterns with wearable smartwatch devices,”International journal of environmental research and public health, vol. 19, no. 3, p. 1279, 2022

    work page 2022

  15. [15]

    A simple field method to identify foot strike pattern during running,

    M. Giandolini, T. Poupard, P. Gimenez, N. Horvais, G. Y . Millet, J.-B. Morin, and P. Samozino, “A simple field method to identify foot strike pattern during running,”Journal of biomechanics, pp. 1588–1593, 2014

    work page 2014

  16. [16]

    Integer-only quantized Transformers for embedded FPGA-based time-series forecasting in AIoT,

    T. Ling, C. Qian, and G. Schiele, “Integer-only quantized Transformers for embedded FPGA-based time-series forecasting in AIoT,” inIEEE Annual Congress on Artificial Intelligence of Things. IEEE, 2024

    work page 2024

  17. [17]

    Tiny time-series Transform- ers: Realtime multi-target sensor inference at the edge,

    T. Becnel, K. Kelly, and P.-E. Gaillardon, “Tiny time-series Transform- ers: Realtime multi-target sensor inference at the edge,” inInternational Conference on Omni-layer Intelligent Systems. IEEE, 2022

    work page 2022

  18. [18]

    Wearable FPGA platform for accelerated DSP and AI applications,

    D. Roggen, R. Cobden, A. Pouryazdan, and M. Zeeshan, “Wearable FPGA platform for accelerated DSP and AI applications,” inInter- national Conference on Pervasive Computing and Communications Workshops and other Affiliated Events. IEEE, 2022, pp. 66–69

    work page 2022

  19. [19]

    ElasticAI: Creating and deploying energy-efficient Deep Learning accelerator for pervasive computing,

    C. Qian, T. Ling, and G. Schiele, “ElasticAI: Creating and deploying energy-efficient Deep Learning accelerator for pervasive computing,” in International Conference on Pervasive Computing and Communications Workshops and other Affiliated Events. IEEE, 2023, pp. 297–299

    work page 2023

  20. [20]

    A hardware and software co-design for energy-efficient Neural Network Accelerator with multiplication-less folded-accumulative PE for radar-based hand gesture recognition,

    F. Li, Y . Guan, and W. Ye, “A hardware and software co-design for energy-efficient Neural Network Accelerator with multiplication-less folded-accumulative PE for radar-based hand gesture recognition,”IEEE Transactions on Very Large Scale Integration (VLSI) Systems, 2024

    work page 2024

  21. [21]

    Eciton: Very low- power Recurrent Neural Network accelerator for real-time inference at the edge,

    J. Chen, S.-W. Jun, S. Hong, W. He, and J. Moon, “Eciton: Very low- power Recurrent Neural Network accelerator for real-time inference at the edge,”ACM Transactions on Reconfigurable Technology and Systems, vol. 17, no. 1, pp. 1–25, 2024

    work page 2024

  22. [22]

    MLoF: Machine Learning accelerators for the low-cost FPGA platforms,

    R. Chen, T. Wu, Y . Zheng, and M. Ling, “MLoF: Machine Learning accelerators for the low-cost FPGA platforms,”Applied sciences, vol. 12, no. 1, p. 89, 2021

    work page 2021

  23. [23]

    Enabling vibration-based gesture recognition on everyday furniture via energy-efficient FPGA implementation of 1D Convolutional Networks,

    K. Shibata, T. Ling, C. Qian, T. Matsui, H. Suwa, K. Yasumoto, and G. Schiele, “Enabling vibration-based gesture recognition on everyday furniture via energy-efficient FPGA implementation of 1D Convolutional Networks,”IEEE Annual Congress on Artificial Intelligence of Things (AIoT), 2025

    work page 2025

  24. [24]

    Automated Energy-Aware Time-Series Model Deployment on Embedded FPGAs for Resilient Combined Sewer Overflow Management

    T. Ling, V . Singh, C. Qian, F. Biessmann, and G. Schiele, “Auto- mated energy-aware time-series model deployment on embedded FPGAs for resilient combined sewer overflow management,”arXiv preprint arXiv:2508.13905, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  25. [25]

    Automating versatile time-series analysis with tiny Transformers on embedded FPGAs,

    T. Ling, C. Qian, L. J. Haßler, and G. Schiele, “Automating versatile time-series analysis with tiny Transformers on embedded FPGAs,” in 2025 IEEE Computer Society Annual Symposium on VLSI (ISVLSI), vol. 1, 2025, pp. 1–6

    work page 2025

  26. [26]

    Optuna: A next- generation hyperparameter optimization framework,

    T. Akiba, S. Sano, T. Yanase, T. Ohta, and M. Koyama, “Optuna: A next- generation hyperparameter optimization framework,” inProceedings of the 25th ACM SIGKDD international conference on knowledge discovery & data mining, 2019, pp. 2623–2631

    work page 2019