EUVIMEDEUVIMED entdecken
Open Access · CC BY-NC-ND 4.0

Improving Human Action Recognition Through Cross-View Camera-Aware Augmentation

Zahid Khaskheli, Qasim Ali, M Shehram Shah, Din Muhammad, Farast Ali, Hina Ali, Maria Khalid, Zoha Waheed, Umer Gaffar

Mehran University Research Journal of Engineering and Technology · 01.07.2026 · 45 (3) · 205-218

DOI: 10.22581/muet1982.0905 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Mehran University Research Journal of Engineering and Technology, online erschienen am 01.07.2026. Verfasst von Zahid Khaskheli und 8 weiteren Autor:innen. Der Volltext ist frei zugänglich (Open Access, CC BY-NC-ND 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Human Action Recognition (HAR) is vital in surveillance, sports analysis, and Human–Computer Interaction (HCI). Publicly available benchmarks mostly suffer from limited sample sizes, class imbalance, occlusions, and viewpoint variability, which may affect model generalization. Most existing HAR augmentation pipelines rely on generic or heuristic transformations that do not explicitly model camera-induced variations. To address these challenges, we propose an Augmented Data Generator Framework (ADGF) for HAR that integrates multiple parameter-controlled augmentation strategies designed to approximate realistic camera-induced variability while preserving temporal consistency. Our method integrates four transformation strategies. Each augmentation technique is parameterized across validated variation ranges, realistically simulating camera viewpoint changes, motion diversity, and partial occlusions, while preserving frame-to-frame temporal coherence, compared with uniformly applied conventional augmentations. ADGF distributes controlled transformations to maintain a balance and preserve action semantics. We evaluate performance with standard metrics, and perform an ablation study to measure each augmentation's contribution. With the combined augmented set, a top-1 accuracy on HMDB51 rises from 87.5% (baseline) to 97.7%, and final training loss falls from 0.21 to 0.12. Required epochs drop from 38 to 25 (≈34% fewer epochs). Cross-dataset testing on UCF101 demonstrates improved robustness, with accuracy increasing from 85.1% to 96.5%. The ablation study shows that each augmentation yields a positive gain with viewpoint variation, that occlusion simulation produces the largest single-technique improvements, and that the full combination produces the strongest transferability. These results indicate that a balanced structured augmentation strategy substantially improves both in-dataset performance and cross-dataset robustness. Implementation details, augmentation configurations, and experimental settings will be made publicly available upon publication to support reproducibility.

Zitieren

APA

Khaskheli, Z., Ali, Q., Shah, M. S., Muhammad, D., Ali, F., Ali, H., Khalid, M., Waheed, Z., & Gaffar, U. (2026). Improving Human Action Recognition Through Cross-View Camera-Aware Augmentation. Mehran University Research Journal of Engineering and Technology, 45(3), 205-218. https://doi.org/10.22581/muet1982.0905

Vancouver

Khaskheli Z, Ali Q, Shah MS, Muhammad D, Ali F, Ali H, et al. Improving Human Action Recognition Through Cross-View Camera-Aware Augmentation. Mehran University Research Journal of Engineering and Technology. 2026;45(3):205-218. doi:10.22581/muet1982.0905

BibTeX
@article{Khaskheli2026,
  author = {Khaskheli, Zahid and Ali, Qasim and Shah, M Shehram and Muhammad, Din and Ali, Farast and Ali, Hina and Khalid, Maria and Waheed, Zoha and Gaffar, Umer},
  title = {Improving Human Action Recognition Through Cross-View Camera-Aware Augmentation},
  journal = {Mehran University Research Journal of Engineering and Technology},
  year = {2026},
  volume = {45},
  number = {3},
  pages = {205--218},
  doi = {10.22581/muet1982.0905}
}
Kostenlos starten →