Drowning incidents in swimming pools remain a critical public health issue globally, where rapid detection and response significantly impact survival rates. Traditional human-based surveillance and sensor-based systems face challenges of cognitive limitations, environmental constraints, and high costs. We propose a two-stage (Two-Stage) framework that, utilizing video streams captured by camera sensors, employs computer vision and deep learning to precisely detect human objects in real time and subsequently classify ‘risk’ behaviors to predict safety incidents. The core focus of this study is (1) to benchmark the performance of the latest real-time object detection models, YOLOv12 and RT-DETR, for the ‘person’ detection module (Stage 1) in a pool environment and (2) to validate the efficacy of a hybrid data strategy—integrating public datasets (Public-Sets) with a custom-collected dataset (Custom-Set)—to optimize this detector’s performance against the known challenge of data scarcity. Experiments were conducted in three scenarios (public data only, custom data only, and a hybrid combination). The results revealed a stark trade-off between detection speed and accuracy; RT-DETR-R50 demonstrated exceptional real-time speeds (approximately 140 FPS), whereas YOLOv12-L provided superior accuracy (mAP) but was not viable for real-time use. We also found that the Public-Set (from Roboflow (9500 images) produced the highest general accuracy (mAP@.5), while the Custom-Set (1,986 images) produced the highest localization precision (mAP@.5:.95). Through this research, an empirical foundation for the ‘detection’ component (Stage 1) of the proposed framework was established and the path for integration with Stage 2 ‘precise behavior classification’ models in future work was outlined.