A Lightweight Drowning Person Detection Using Deep Learning Algorithm

Abstract

Visual obstructions and fatigue often hinder human surveillance in preventing drowning incidents. Manual surveillance methods are susceptible to visual obstructions, distractions from crowds, and fatigue. To address these issues, this study proposes an automated real-time drowning detection system that utilizes state-of-the-art deep learning techniques. We use the YOLOv12-Nano architecture, selected for its balance between detection accuracy and computational efficiency. This model was trained and evaluated on the SelfMade dataset, which covers various water conditions and poses indicating swimmers in distress. In testing, YOLOv12-Nano achieved a mAP@50 of 0.984 and a mAP@50-95 of 0.732, with 2.52 million parameters and a computational requirement of 6 GFLOPs. These results demonstrate that YOLOv12-Nano-based automatic detection provides reliable, resource-efficient real-time monitoring, is suitable for implementation on real-world application, and can support human surveillance and accelerate emergency responses to reduce fatal drowning accidents.