An Indoor Pool Drowning Risk Detection Method Based on Improved YOLOv4
Qiang Niu, Yu-Cheng Wang, Songhe Yuan, Kaixia Li, Xiaohu Wang
Abstract
Drowning accidents in indoor swimming pools are growing in recent years, especially for children. Traditional manual inspection by a lifeguard is costly in economics, which generates the demand for a much more intelligent AI system to play an important role in swimmer safety. In this paper, we propose a new drowning risk detection method based on YOLOv4 with a MA_CBAM module to build an effective drowning early-warning system. The proposed method, MA_CBAM-YOLOv4, inherits the basic idea of the YOLOv4, but uses the Meta-ACON activation function to replace the regular “Relu” activation function of the Neck part of YOLOV4 and adds the CBAM module with the Meta-ACON activation function. The proposed method was tested on a constructed swimming video dataset. The experimental results showed that the improved model MA_CBAM-YOLOv4 performs well with higher accuracy and robustness compared with the original yolov4. The improved method has higher evaluation indexes for each category and reaches a mean average precision (MAP) of 86.92%, which exceeds the original method by 1.82%.