A Comparative Study on the Performance of Object Detection Model for Real-Time Safety Accident Detection in Swimming Pool

This study presents a real-time object detection-based safety management framework designed for recognizing dangerous behaviors in swimming pool environments. A custom dataset was constructed using images that reflect real-world swimming pool scenarios and annotated across five action categories: Swimming, Normal, Slipping, Drowning, and Floundering. To address data imbalance, especially in rare but critical behaviors such as Drowning and Floundering, class-specific augmentation techniques were applied. The performance of two representative model families was compared: the CNN-based YOLO series (YOLOv9, YOLOv10, YOLOv11) and the Transformer-based RT-DETR series (RT-DETR, RT-DETRv2). Experimental results demonstrate that YOLOv11L achieved the best balance between accuracy, inference speed, and computational efficiency, making it a suitable candidate for real-time safety monitoring systems. Although the RT-DETR models exhibited strong precision and recall, their higher computational cost and lower FPS make them less favorable in latency-sensitive scenarios. Furthermore, while the Slipping class was derived from simulated images, it did not negatively impact the detection performance of other classes. Future work will focus on improving the generalization of the detection system by incorporating real-world Slipping data and evaluating performance across diverse swimming pool conditions.