Drowning is one of the leading global causes of unintentional injury-related deaths, especially among children, highlighting the critical need for innovative monitoring solutions. Traditional surveillance methods often fail due to environmental factors such as glare, motion artifacts, and human error. To address these challenges, this paper introduces a manually curated and annotated underwater dataset of 69,512 frames, capturing diverse aquatic scenarios, including normal swimming, struggling behaviors, and simulated drowning. Using this dataset, we developed and evaluated a real-time drowning detection model based on YOLOv8n, specifically optimized to handle underwater-specific challenges such as optical distortion, lighting variability, and occlusion. The system supports multi-swimmer detection in crowded aquatic environments and achieves robust performance under varied conditions. Experimental results demonstrate that the proposed model achieves 98.3% precision with real-time inference of 22 ms per frame (45 FPS). These results set a new benchmark for underwater safety systems and provide a strong foundation for future research in aquatic computer vision applications.