A Novel Drowning Detection System and Its Performance Analysis
1
Issued Date
2025-01-01
Resource Type
Scopus ID
2-s2.0-105014408312
Journal Title
22nd International Conference on Electrical Engineering Electronics Computer Telecommunications and Information Technology Ecti Con 2025
Rights Holder(s)
SCOPUS
Bibliographic Citation
22nd International Conference on Electrical Engineering Electronics Computer Telecommunications and Information Technology Ecti Con 2025 (2025)
Suggested Citation
Iampanit T., Somprakij P., Siriwanna N., Iampanit N., Jirakittayakorn N. A Novel Drowning Detection System and Its Performance Analysis. 22nd International Conference on Electrical Engineering Electronics Computer Telecommunications and Information Technology Ecti Con 2025 (2025). doi:10.1109/ECTI-CON64996.2025.11100770 Retrieved from: https://repository.li.mahidol.ac.th/handle/123456789/111946
Title
A Novel Drowning Detection System and Its Performance Analysis
Corresponding Author(s)
Other Contributor(s)
Abstract
Drowning is one of the leading causes of accidental death worldwide. It often occurs silently within seconds. Although lifeguards and surveillance systems are available, human oversight can also be erroneous due to fatigue, poor visibility, and crowded conditions. To alleviate the problem, this study proposes a drowning detection system that processes high-resolution (4K) video streams. The proposed system consists of four key components: (1) Object Detection using You Only Look Once (YOLO) version 11 to detect individuals, (2) Tracking Module to maintain consistent identities across the frames, (3) Image Classification using ResNet50 as a feature extractor to distinguish between drowning and normal swimming on a frame-by-frame basis, and (4) Grid Zone Monitoring to verify consecutive detections within defined spatial regions for reducing false alarms. A custom-made swimming pool dataset was utilized to evaluate the performance of the developed system on image-level and event-level. Additionally, multiple feature extractors were used as comparators. The results found that using ResNet50 as the feature extractor achieved the highest performance with an accuracy of 91.84% and recall of 80.82% on image-level. It attained an accuracy of 100.00% on event-level, demonstrating its effectiveness in continuous monitoring.