EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

Improving Crowded Head Detection With Task-Aware Focal Modulation and Feature-Crossing Knowledge Distillation

Yuteng Sui, Xinaghua Shan, Linlin Dai, Hui Jing, Hualong Sun, Jianjun Ma

IEEE Access · 2026 · 14 · 69414-69425

DOI: 10.1109/access.2026.3690800 ↗

Kurz gefasst

Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Yuteng Sui und 5 weiteren Autor:innen. Die Arbeit nennt 34 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: China Academy of Railway Sciences Corporation Ltd..

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Accurate head detection in crowded scenes is a fundamental requirement for intelligent surveillance systems, yet it remains challenging due to the prevalence of tiny targets, severe occlusion, and strict efficiency constraints in real-world deployment. To address these issues, this paper proposes TAFM-KD (Task-Aware Focal Modulation Knowledge Distillation) for lightweight head detection in dense crowd scenarios. The proposed framework employs a feature-crossing prediction alignment strategy that evaluates student feature representations in the teacher’s prediction space, thereby mitigating supervision conflicts between hard ground-truth labels and soft teacher outputs. Furthermore, a task-aware focal modulation mechanism is introduced to emphasize poorly localized head instances during training. By coupling these strategies, our method provides a shared feature space to prioritize hard samples for effectively distilling knowledge from teacher to student models, thereby improving the overall detection performance. Compared with representative compact one-stage detectors, extensive experiments demonstrate that the proposed approach achieves superior accuracy in densely crowded scenarios, while maintaining real-time inference speed with only 1.62M parameters. These results indicate that the proposed framework provides an effective accuracy–efficiency trade-off for practical head detection in crowded surveillance applications.

Förderung

  • China Academy of Railway Sciences Corporation Ltd.

Zitieren

APA

Sui, Y., Shan, X., Dai, L., Jing, H., Sun, H., & Ma, J. (2026). Improving Crowded Head Detection With Task-Aware Focal Modulation and Feature-Crossing Knowledge Distillation. IEEE Access, 14, 69414-69425. https://doi.org/10.1109/access.2026.3690800

Vancouver

Sui Y, Shan X, Dai L, Jing H, Sun H, Ma J. Improving Crowded Head Detection With Task-Aware Focal Modulation and Feature-Crossing Knowledge Distillation. IEEE Access. 2026;14:69414-69425. doi:10.1109/access.2026.3690800

BibTeX
@article{Sui2026,
  author = {Sui, Yuteng and Shan, Xinaghua and Dai, Linlin and Jing, Hui and Sun, Hualong and Ma, Jianjun},
  title = {Improving Crowded Head Detection With Task-Aware Focal Modulation and Feature-Crossing Knowledge Distillation},
  journal = {IEEE Access},
  year = {2026},
  volume = {14},
  pages = {69414--69425},
  doi = {10.1109/access.2026.3690800}
}
Kostenlos starten →