Implementation of deep-learning-based edge computing for preventing drowning

With the increasing popularity of computer vision and ocean engineering, maritime visual surveillance has emerged as an important area of study. Despite the development of various detection or surveillance techniques for maritime locations, maritime visual surveillance remains a challenge owing to the complex, unconstrained, and diverse nature of such places. In addition, only a few studies have investigated edge computing in actively preventing people from drowning. Considering several people lose their lives due to drowning, the flourishing technologies of artificial intelligence (AI) and deep learning have been used for our study. In this paper, an implementation of deep-learning-based edge computing to prevent drowning with the use of NVIDIA Jetson Nano is proposed.


Introduction
According to the global report of World Health Organization (WHO) (1) , drowning is the 3rd leading cause for unintentional deaths worldwide, accounting for 7 % of all injury-related deaths. More than 360,000 people drown worldwide each year. Thus, drowning is a major public health concern (1,2). Notably, over half of the world's drowning cases occur in the WHO's Western Pacific Region and WHO's South-East Asia Region.
Despite the development of various surveillance for object detection or recognition in maritime locations, maritime visual surveillance remains a challenge owing to the complex, unconstrained, and diverse nature of such environments. Electro-optical (EO) sensors are typically used for security-related visual surveillance purposes, such as ship tracking, counting, and classification (3,4,5,6) .
However, as 24-hour manual surveillance is impractical, achieving automatic maritime surveillance using object and foreground detection has become an important goal. In addition, the development of video surveillance applications, for utilization in maritime environments, has been hindered by the noise and complexity of sea scenes that arise from factors such as water motion, dynamic backgrounds, waves, sea foam, water spray, reflections, wave ripples, and wake (5,7) . Furthermore, because harbors comprise a vast area and are not patrolled or monitored all the time even if it is a naval port, the crucial period to rescue a drowning person is overlooked. Although To overcome these problems, this paper proposes an active protection edge computing device that can ensure safety through the use of a simple visual camera, which provides early warnings when a potential drowning victim is noticed and alarms the security guards. In this paper, the implementation of deep-learning-based edge computing to prevent drowning with the use of NVIDIA Jetson Nano is proposed.

Methods
The flowchart of the proposed method is shown in Fig

Dataset Collection and Pre-processing
For collecting the required drowning image dataset, 30 volunteers, unrelated to the study, were asked to enter the water, and they were made to wear three kinds of costumes, which were the navy uniform, navy fatigue dress, and swimsuit, as shown in Fig. 2.
These volunteers were asked to make different poses, as shown in Fig  During the pre-processing step, each image was normalized to 224×224 resolution.  The training phase flowchart is illustrated in Fig. 6.
Firstly, the collected drowning dataset is randomly 236 Proceedings of the 8th IIAE International Conference on Industrial Application Engineering 2020 partitioned into three-fourth for the training model, as illustrated in Fig. 6(a) and one-fourth for testing model. Firstly, the samples are sent to a computer, as illustrated in Fig. 6(a). Secondly, the code is programmed in the Jupyter Notebook, as illustrated in Fig. 6(b), to create a model for identifying drowning and not-drowning images. Thirdly, as illustrated in Fig.   6(c), a model is deployed to Jetson Nano, which can be combined with the camera.   In the drowning prevention device, we mounted the Raspberry Pi Camera, whose model product was Pi NoIR Camera V2, on NVIDIA Jetson Nano. Further, the training deep learning model was deployed in Jetson Nano and evaluated in a real environment.

Experimental Analysis
The implemented drowning detection device, as shown in Fig. 9(a) Fig. 10(b). We note that the prediction results were not as stable as expected.
Specifically, when someone was drowning, the device might alter the prediction leading to inaccurate determination of whether the person was drowning or not.  3) Lack of spatiotemporal clues: In the future, we will add new images in the drowning dataset using the Raspberry Pi Camera. In addition, data augmentation will also be applied to create a robust training model that can adapt to different sensor sources, shot angles, and illumination conditions.
Furthermore, we will improve our device by using object detection in video sequences.

Conclusions
In this study, we implemented a deep-learning-based edge computing device for prevention of drowning. In addition, it was shown that edge computing with AI technology can be realistically applied in our lives. Apart from mounting our drowning prevention device on the shore, it also can be deployed in an autonomous lifebuoy, which has propeller blades and can move automatically, as an active-preventing drowning system. This would reduce the complexity involved in saving people from drowning and greatly increase people's safety both in the harbor and shore areas.