A video system based on convolutional autoencoder for drowning detection

Computer vision combined with deep learning technologies is widely used in video surveillance. In this paper, it is applied to drowning detection video systems. Traditional drowning detection methods detect drowning mainly by monitoring the physiological condition, the time and motion of swimmers in the water. But these methods are not applicable to detect early quiet drowning phenomena. Some researchers realize supervised classification by simulative drowning features. But real drowning events are difficult to truly simulate, so these methods are not reliable. In this paper, a drowning detection video system with edge computing is proposed, and it can detect drowning events in swimming pools without any wearable devices. According to the characteristics of drowning people, the strategies for underwater near-vertical human detection are proposed, providing a reliable basis for drowning detection. A lightweight drowning detection convolutional autoencoder is proposed to achieve unsupervised drowning detection, solving the lack of drowning videos and the inauthenticity of simulative videos. Then, an edge device is designed for detecting drowning in real time at the edge. Finally, for training and experimental evaluation, a pool dataset including many pool underwater video sequences is produced. The experimental results show that the proposed drowning detection method has a good comprehensive performance. The system is feasible and valuable.

Automated summary

In this paper, a drowning detection video system that combines computer vision and deep learning technologies is proposed. It can detect drowning events in swimming pools in real time, solving the problem of inaccurate drowning detection in traditional methods. By proposing strategies for underwater near-vertical human detection and a lightweight drowning detection convolutional autoencoder, the lack of drowning videos and the inauthenticity of simulative videos are addressed. Experimental results demonstrate that the proposed method has good comprehensive performance.