Development of outdoor swimmers detection system with small object detection method based on deep learning

Wild swimming, or swimming in prohibited outdoor places, is a major source of drowning occurrences and a key problem in outdoor water safety management. Currently, manual patrol and warning signs are the basic methods adopted by the local government for outdoor water safety management to inspect drowning accidents. However, they are inefficient, costly, and of little avail. To this goal, a novel object detector for outdoor swimmers was developed via transfer learning utilizing the Microsoft Common Objects in Context (MS COCO) dataset as a training starting point. The model was then evaluated and retrained to possess the capacity to classify swimmers, suspected swimmers, and pedestrians. The total precision and detection time of our proposed swimmer detection with small object detection approach are 99.45% and 43.44 ms, respectively, which are greater than those of existing methods and traditional data augmentation methods. We verified the effectiveness of the proposed method on small target detection and designed two prototypes of hardware systems (fixed monitoring device and drone monitoring device) to meet the requirements of stationary and movable detection scenarios that can identify and warn of the possible phenomenon of wild swimming efficiently. This scheme can provide a more comprehensive reference for other innovative city applications that rely on cameras and can be valuable for society.