Lightweight multi-task network for synchronous fatigue detection and ankle instability gait analysis in football training with edge deployment
- 1Qingyuan Polytechnic
Abstract
Football training imposes considerable mechanical stress on lower extremities, which makes real-time monitoring of athlete fatigue and ankle joint health essential for injury prevention. Existing approaches treat fatigue detection and abnormal gait analysis as independent tasks, producing redundant computation and overlooking shared biomechanical signatures between these two phenomena. This paper proposes a lightweight multi-task network that performs, in a single forward pass, fatigue detection and ankle instability gait analysis from monocular football training video, with explicit optimization for edge-device deployment. The architecture employs a shared depthwise-separable convolutional backbone enhanced with coordinate attention, feeding two compact task-specific heads through a multi-scale feature fusion module. A homoscedastic uncertainty-weighted loss balancing strategy with staged warm-up scheduling addresses gradient dominance between tasks of differing convergence rates. Experiments on a dataset of 4,860 annotated clips from 42 male collegiate athletes demonstrate that the multi-task formulation reaches 92.3% fatigue detection accuracy and 87.6% gait instability classification F1 with an ankle inversion-angle regression MAE of 2.74°. Improvements over single-task baselines are statistically significant (paired Wilcoxon test, p < 0.01) while operating at only 0.44 G FLOPs. Quantitative robustness analyses across ankle keypoint occlusion, illumination, and motion-blur conditions further delineate the practical degradation envelope. INT8 quantization lowers inference latency to 8.6 ms on the Jetson Orin NX at a minimal accuracy cost, which indicates that real-time operation is achievable on resource-constrained edge hardware. We note that the present cohort is restricted to male collegiate athletes from a single institution and that the structured drill setting differs from the intermittent, high-acceleration profile of match play; external validation across broader populations and competitive scenarios remains an essential next step.
Related
- Comparison of lower extremity peak muscle forces and force-generating capacities during the functional side hop test in male collegiate football players with and without chronic ankle instability
- Lightweight multimodal brain tumor MRI segmentation network for efficient clinical deployment: MLG-UNet achieves an average Dice of 86.67% with 0.81 M parameters
- Multi-Version Evaluation of Deep Learning Architectures for Tibial Plateau Fracture Detection and Deployment in a Web-Based Clinical Support System.
- Image-Based Breast Implant Detection for Mammography Dataset Curation and Near-Real-Time Deployment: Comparing Foundation Models and Task-Specific Convolutional Models
- A lightweight deep learning model for real-time in-vehicle driver distraction detection with low-latency inference
BibTeX
@article{maoning2026lightweight,
title = {{Lightweight multi-task network for synchronous fatigue detection and ankle instability gait analysis in football training with edge deployment}},
author = {Zou Maoning},
journal = {Scientific Reports},
year = {2026},
month = aug,
doi = {10.1038/s41598-026-65579-1},
url = {https://doi.org/10.1038/s41598-026-65579-1},
note = {CC BY},
}
RIS
TY - JOUR
TI - Lightweight multi-task network for synchronous fatigue detection and ankle instability gait analysis in football training with edge deployment
AU - Zou Maoning
PY - 2026
DA - 2026/08/01
JO - Scientific Reports
DO - 10.1038/s41598-026-65579-1
UR - https://doi.org/10.1038/s41598-026-65579-1
AB - Football training imposes considerable mechanical stress on lower extremities, which makes real-time monitoring of athlete fatigue and ankle joint health essential for injury prevention. Existing approaches treat fatigue detection and abnormal gait analysis as independent tasks, producing redundant computation and overlooking shared biomechanical signatures between these two phenomena. This paper proposes a lightweight multi-task network that performs, in a single forward pass, fatigue detection and ankle instability gait analysis from monocular football training video, with explicit optimization for edge-device deployment. The architecture employs a shared depthwise-separable convolutional backbone enhanced with coordinate attention, feeding two compact task-specific heads through a multi-scale feature fusion module. A homoscedastic uncertainty-weighted loss balancing strategy with staged warm-up scheduling addresses gradient dominance between tasks of differing convergence rates. Experiments on a dataset of 4,860 annotated clips from 42 male collegiate athletes demonstrate that the multi-task formulation reaches 92.3% fatigue detection accuracy and 87.6% gait instability classification F1 with an ankle inversion-angle regression MAE of 2.74°. Improvements over single-task baselines are statistically significant (paired Wilcoxon test, p < 0.01) while operating at only 0.44 G FLOPs. Quantitative robustness analyses across ankle keypoint occlusion, illumination, and motion-blur conditions further delineate the practical degradation envelope. INT8 quantization lowers inference latency to 8.6 ms on the Jetson Orin NX at a minimal accuracy cost, which indicates that real-time operation is achievable on resource-constrained edge hardware. We note that the present cohort is restricted to male collegiate athletes from a single institution and that the structured drill setting differs from the intermittent, high-acceleration profile of match play; external validation across broader populations and competitive scenarios remains an essential next step.
SN - 2045-2322
ER -