Swimming-YOLO: a drowning detection method in multi-swimming scenarios based on improved YOLO algorithm

Drowning is now a global public safety issue, and there is a significant demand for the urgent detection and warning of drowning incidents. Some neural network-based object detection algorithms have been proposed to promptly identify and locate drowning individuals, which have improved the chances of survival to a certain degree. However, these algorithms are still limited by the water environment and there exists substantial scope for enhancing the precision of detection. To improve the accuracy of detecting drowning persons in complex swimming scenarios, this study proposes swimming-YOLO, a drowning object detection model. Firstly, deformable convolution is used to improve the model, which shifts the sampling points of convolution to more salient locations, making it easier for the model to distinguish between the background and the drowning person. Secondly, deformable attention is introduced into the model, allowing the attention module to focus on relevant regions while enhancing the extraction of crucial detail features. Such enhancement helps the model more effectively distinguish between swimmers and drowning individuals. Finally, this study introduces auxiliary detection heads and uses InnerIOU to modify the loss function during the training process. Such addition provides the model with more comprehensive information during the training and improves its generalization ability on drowning datasets. In the experiment, a dataset dedicated to drowning detection is collected from multiple drowning scenarios. The results show that swimming-YOLO owns the best overall detection accuracy and drowning detection accuracy, and its speed meets practical requirements.