Multi-Head Attention-Based Framework with Residual Network for Human Action Recognition
Sensors · 2025
DOI: 10.3390/s25092930 ↗Kurz gefasst
Fachartikel in der Zeitschrift Sensors, 2025. Verfasst von Basheer Al-Tawil und 3 weiteren Autor:innen. Beteiligte Einrichtung: Neuro-Information Technology, Otto-von-Guericke-University Magdeburg, 39106 Magdeburg, Germany. Hauptthemen laut MEDLINE: Pattern Recognition, Automated, Attention. Die Arbeit wurde bisher 7-mal zitiert (Crossref) und nennt 61 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: Federal Ministry of Education and Research of 378 Germany (BMBF), German Research Foundation 379 (DFG- Roboter SEMIAC), European Regional Develop- 380 ment Fund (EFRE-ENABLING).
Abstract (Originalsprache)
Human action recognition (HAR) is essential for understanding and classifying human movements. It is widely used in real-life applications such as human–computer interaction and assistive robotics. However, recognizing patterns across different temporal scales remains challenging. Traditional methods struggle with complex timing patterns, intra-class variability, and inter-class similarities, leading to misclassifications. In this paper, we propose a deep learning framework for efficient and robust HAR. It integrates residual networks (ResNet-18) for spatial feature extraction and Bi-LSTM for temporal feature extraction. A multi-head attention mechanism enhances the prioritization of crucial motion details. Additionally, we introduce a motion-based frame selection strategy utilizing optical flow to reduce redundancy and enhance efficiency. This ensures accurate, real-time recognition of both simple and complex actions. We evaluate the framework on the UCF-101 dataset, achieving a 96.60% accuracy, demonstrating competitive performance against state-of-the-art approaches. Moreover, the framework operates at 222 frames per second (FPS), achieving an optimal balance between recognition performance and computational efficiency. The proposed framework was also deployed and tested on a mobile service robot, TIAGo, validating its real-time applicability in real-world scenarios. It effectively models human actions while minimizing frame dependency, making it well-suited for real-time applications.
Schlagwörter
MeSH-Schlagwörter (MEDLINE)
Stichwörter der Autor:innen
Autor:innen und Institutionen
- Neuro-Information Technology, Otto-von-Guericke-University Magdeburg, 39106 Magdeburg, Germany.
ORCID-Profile: Basheer Al-Tawil, Magnus Jung, Thorsten Hempel, Ayoub Al-Hamadi
Förderung
- Federal Ministry of Education and Research of 378 Germany (BMBF)
- German Research Foundation 379 (DFG- Roboter SEMIAC)
- European Regional Develop- 380 ment Fund (EFRE-ENABLING)
- EFRE-RoboLab
- Federal Ministry of Education and Research of Germany (BMBF)
- German Research Foundation
- European Regional Development Fund
Zitieren
APA
Al-Tawil, B., Jung, M., Hempel, T., & Al-Hamadi, A. (2025). Multi-Head Attention-Based Framework with Residual Network for Human Action Recognition. Sensors. https://doi.org/10.3390/s25092930
Vancouver
Al-Tawil B, Jung M, Hempel T, Al-Hamadi A. Multi-Head Attention-Based Framework with Residual Network for Human Action Recognition. Sensors. 2025. doi:10.3390/s25092930
BibTeX
@article{AlTawil2025,
author = {Al-Tawil, Basheer and Jung, Magnus and Hempel, Thorsten and Al-Hamadi, Ayoub},
title = {Multi-Head Attention-Based Framework with Residual Network for Human Action Recognition},
journal = {Sensors},
year = {2025},
doi = {10.3390/s25092930}
}
EUVIMED