EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

Biomechanics-Informed Masked Autoencoder With Anchor Alignment for Robust Human Activity Recognition

Ting Guo, Feng Hou, Yan Pang, Zhongwei Wang

IEEE Access · 2026 · 14 · 106955-106969

DOI: 10.1109/access.2026.3713236 ↗

Kurz gefasst

Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Ting Guo und 3 weiteren Autor:innen. Die Arbeit nennt 56 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: National Social Science Fund Project, Key Research and Development Program of Hunan Province of China, Hunan Key Laboratory of Intelligent Logistics Technology.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Wearable Human Activity Recognition (HAR) is critical for occupational safety in dynamic logistics environments, where falls and musculoskeletal injuries impose substantial economic and workforce burdens. In particular, this work focuses on balance-critical activity classification, where the reliable discrimination of posturally similar movements—such as walking, stumbling, and load-bearing transitions—is essential for early fall risk detection and injury prevention. Despite recent advances in self-supervised learning (SSL), existing HAR models largely neglect biomechanical principles, rendering them vulnerable to performance degradation under sensor placement shifts, modality heterogeneity, and movement variability. This study develops a robust HAR framework that overcomes the instability and poor generalization of current SSL methods by explicitly integrating human biomechanics with physics-informed anchor-guided feature alignment. We propose BIMA (Biomechanics-Informed Masked Autoencoder with Anchor Alignment), which integrates biomechanical priors, masked self-supervised pretraining, and anchor-guided representation learning. BIMA comprises three synergistic components: a biomechanical balance-aware sampling and encoding module; a sensor-wise masked pretraining module; and an anchor-guided feature fusion module. BIMA achieves state-of-the-art performance while substantially reducing computational overhead and minimizing misclassification among activity-similar categories. By transitioning HAR to a physics-informed paradigm, BIMA provides principled guidance for occupational safety monitoring in resource-constrained industrial settings.

Förderung

  • National Social Science Fund Project
  • Key Research and Development Program of Hunan Province of China
  • Hunan Key Laboratory of Intelligent Logistics Technology

Zitieren

APA

Guo, T., Hou, F., Pang, Y., & Wang, Z. (2026). Biomechanics-Informed Masked Autoencoder With Anchor Alignment for Robust Human Activity Recognition. IEEE Access, 14, 106955-106969. https://doi.org/10.1109/access.2026.3713236

Vancouver

Guo T, Hou F, Pang Y, Wang Z. Biomechanics-Informed Masked Autoencoder With Anchor Alignment for Robust Human Activity Recognition. IEEE Access. 2026;14:106955-106969. doi:10.1109/access.2026.3713236

BibTeX
@article{Guo2026,
  author = {Guo, Ting and Hou, Feng and Pang, Yan and Wang, Zhongwei},
  title = {Biomechanics-Informed Masked Autoencoder With Anchor Alignment for Robust Human Activity Recognition},
  journal = {IEEE Access},
  year = {2026},
  volume = {14},
  pages = {106955--106969},
  doi = {10.1109/access.2026.3713236}
}
Kostenlos starten →