<![CDATA[Although effective in swimming pools, automated drowning surveillance in open-water recreational areas remains challenging because of dynamic conditions such as water reflections, lighting variations, wave motion, and crowded scenes. This study proposes the YOLO-NAS-SO (You Only Look Once – Neural Architecture Search – Spatial Optimization) framework to improve the spatial feature representation of YOLO-NAS while maintaining real-time computational efficiency. The framework integrates a Spatial Attention Module after the backbone feature extraction stage to emphasize relevant object features and reduce background interference. For continuous safety monitoring, the proposed framework combines a Modified ByteTrack pipeline with a temporal drowning-duration estimation module. The module monitors the persistence of drowning-risk events across consecutive video frames and generates an early warning when the event duration exceeds a predefined temporal threshold. Unlike conventional ByteTrack, the proposed extension retains the original data association mechanism while adding temporal risk assessment to the tracking process. The main contribution of this study is the integration of spatial attention-based feature enhancement with temporal risk assessment for drowning surveillance in complex open-water environments. Experimental results show that YOLO-NAS-SO achieves a mAP@50–95 of 0.7075, mAP@50 of 0.9166, precision of 0.8844, recall of 0.8817, and F1-score of 0.8830, outperforming the baseline YOLO-NAS and slightly exceeding YOLOv8 while maintaining a real-time inference time of 1.3 ms. These results indicate that combining spatial feature enhancement with temporal drowning-duration monitoring improves detection robustness and supports continuous early-warning surveillance. The proposed framework therefore provides a practical approach for intelligent drowning-risk monitoring in open-water recreational environments.]]>
Copyrights © 2026
