Author(s): SIEBOLD, E., SCHWEIZER, T., LIZARRIBAR CARRILLO, Y., GILGEN-AMMANN, R., Institution: SWISS FEDERAL INSTITUTE OF SPORT MAGGLINGEN, Country: SWITZERLAND, Abstract-ID: 366

INTRODUCTION:
Long-distance walking is common in military settings and associated with high physical strain and injury risk.[1] Robust gait recognition could improve monitoring of cumulative load in soldiers during prolonged walking periods. Gait recognition can achieve high accuracy but performance depends on sensor placement, speed, and environment.[2,3] This study aims to investigate the laboratory to field transfer of gait recognition models using upper-arm worn inertial measurement units (IMUs) and applying two distinct machine learning approaches.
METHODS:
Laboratory protocols included treadmill walking and running, cycling, strength and high-intensity exercises, sitting, and lying. Field data were obtained during a 90 km military march including road marching and hiking terrain. Triaxial acceleration was recorded at 52 Hz using a wearable IMU. Binary activity labels (gait / non-gait) were derived from experimental protocols in the laboratory and from GPS speed in the field data. Acceleration magnitude was segmented into 5 s windows. Two machine learning approaches were applied: (1) a convolutional neural network (CNN) using z-normalized acceleration magnitude, and (2) a frequency-domain model using fast Fourier transformation features with XGBoost. Models were trained on laboratory data and tested on laboratory and field data to assess transferability, followed by training with combined laboratory and field data and separate testing for each condition. Each model pair was data matched and balanced for participants, condition, and label. Five-fold cross-validation and 50-fold randomized hyperparameter search were applied. Performance was evaluated using F1-score. The Python code is publicly available at Zenodo (DOI: 10.5281/zenodo.18271909).
RESULTS:
19 participants (m = 10, f = 9, age = 32.4 +/- 8.5) completed the laboratory protocol and 24 (m = 23, f = 1, age = 21.4 +/- 1.6) contributed field data, yielding approximately 3.8*10^4 and 2.9*10^5 labeled windows, respectively. When trained on laboratory data, the CNN achieved an F1-score of 0.994 +/- 0.005 on laboratory data and 0.992 +/- 0.004 on field data. The frequency-based model showed comparable laboratory performance (0.995 +/- 0.002) but slightly lower field performance (0.960 +/- 0.035). Including field data in training improved field performance of the frequency-based model to 0.990 +/- 0.003.
CONCLUSION:
Both raw-signal and frequency-domain models achieve high accuracy for gait versus non-gait classification. However, frequency-based models trained on laboratory data show slightly reduced transferability to field settings involving road marching and hiking, whereas raw magnitude-based CNNs are more robust. All models performed similar or better compared to recently proposed algorithms,[2,3] supporting their applicability in real-world monitoring using upper-arm acceleration.

[1] Lovalekar et al., J Sci Med Sport, 2021 [2] Davis et al., Physiol Meas, 2021 [3] Straczkiewicz et al., npj Digit Med, 2023