EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

Weakly supervised learning for drowning detection in over-water construction from videos

Wenkang Guo, Gangyan Xu, Changsheng Qu, Heng Li, Haosen Chen, Lei Hou, Guomin Zhang, Yushu Yang

Journal of Building Design and Environment · 08.05.2026

DOI: 10.70401/jbde.2026.0039 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Journal of Building Design and Environment, online erschienen am 08.05.2026. Verfasst von Wenkang Guo und 7 weiteren Autor:innen. Die Arbeit nennt 26 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Open-water drowning is a leading serious-injury/fatality risk at public waterfronts and in construction over or adjacent to water, where long stand-off views, glare, waves, and occlusions hinder timely detection. We propose TimeSformer+MIL, a weakly supervised temporal framework for event-level drowning monitoring designed for deployment in safety-critical construction settings and aligned with supervisory workflows. The system standardizes video streams into short clips, extracts spatiotemporal evidence with a divided space–time TimeSformer, and aggregates clip scores via top-k multiple-instance learning with a lightweight consistency prior to stabilize weak labels and support calibration. By avoiding person detection and multi-object tracking, the pipeline reduces engineering complexity and failure modes common in cluttered, low-light, or small-scale scenes, improving reliability without increasing operator load. We curate an open-water dataset spanning construction and public-waterfront contexts and evaluate with event-focused metrics aligned to risk governance: recall at target false-positives per hour, alert latency relative to rescue windows, and calibration-aware ranking and thresholded decisions for alerting and escalation. Across clip lengths and aggregation strategies, the approach delivers robust discrimination and translates it into stable, low-latency alerts that meet rescue-time targets while limiting alarm fatigue. For architectural practice and construction safety management, the framework offers a practical path to augment human surveillance with machine attention, functioning as a verifiable administrative control within the hierarchy of controls and integrating with site safety processes to accelerate incident recognition and strengthen risk governance in dynamic open-water settings.

Zitieren

APA

Guo, W., Xu, G., Qu, C., Li, H., Chen, H., Hou, L., Zhang, G., & Yang, Y. (2026). Weakly supervised learning for drowning detection in over-water construction from videos. Journal of Building Design and Environment. https://doi.org/10.70401/jbde.2026.0039

Vancouver

Guo W, Xu G, Qu C, Li H, Chen H, Hou L, et al. Weakly supervised learning for drowning detection in over-water construction from videos. Journal of Building Design and Environment. 2026. doi:10.70401/jbde.2026.0039

BibTeX
@article{Guo2026,
  author = {Guo, Wenkang and Xu, Gangyan and Qu, Changsheng and Li, Heng and Chen, Haosen and Hou, Lei and Zhang, Guomin and Yang, Yushu},
  title = {Weakly supervised learning for drowning detection in over-water construction from videos},
  journal = {Journal of Building Design and Environment},
  year = {2026},
  doi = {10.70401/jbde.2026.0039}
}
Kostenlos starten →