Marine drowning detection method based on improved YOLOv5

In this study, we developed a drowning detection method named AquaYOLO, which enhances the YOLOv5 framework by integrating the detection head mechanism from YOLOX. This involves decoupling the classification and regression prediction heads and incorporating a Dual Attention Net module that combines Channel Attention with Position Attention. Additionally, we added a small target detection layer with 4x4 pixel resolution. To improve border positioning accuracy, we replaced the conventional IoU loss function with the SIoU loss function. Our experiments demonstrate that AquaYOLO achieves a detection accuracy of 92.24% and a recall rate of 82.612% on the dataset, which is 9.98% higher than the traditional YOLOv5s. When compared to YOLOv3, YOLOv4, and Faster CNN, AquaYOLO shows superior detection accuracy, indicating its effectiveness in drowning detection scenarios. This study offers a significant advancement in marine drowning rescue solutions.