High accuracy intelligent real-time framework for detecting infant drowning based on deep learning

Timely detection of swimming accidents plays a critical role in keeping the safety of infants in pools. However, the targets in the infant drowning detection are generally small, densely-distributed and often immersed in complicated environment with lighting various and toy disturbances, and thus there are currently few efficient, accurate and comprehensive solutions for such a task. Aiming at this problem, a novel intelligent real-time framework for detecting infant drowning is proposed. Processing of the framework includes two stages: in stage I, an attention mechanism-integrated YOLOv5-alike model is designed to extract infant portrait foreground, which can reduce the interferences caused by background noise and improve the accuracy of detecting tiny objects. On this basis, in stage II, Single Shot Multi-box Detector (SSD) is utilized to filter out the falsely-detected targets in stage I and recognize swimming posture accurately. Live videos are collected from ten infant swimming pools to generate training dataset of totally 7,296 swimming images and 7,723 portrait images of infants for stage I and II respectively. And a set of 1,222 infant swimming images from another two pools are annotated to test the trained framework. Experimental results show that the Mean Average Precision (mAP) of the framework is 97.17%, and the processing speed can reach 43 frames per second, which outperforms any previous related network and has significant application value in practical infant drowning detection. Additionally, a web-based platform is developed to put the framework into practice, whose preliminary test results demonstrate its strong application potential.