INO-YOLO: A Lightweight Real-Time Deep Learning Model for Drowning Detection on Low-Cost Edge Devices
Applied Technology- Grade:
- 11
- Teacher:
- Kevin Nieves Pichardo
Drowning remains a serious public health issue, causing 230,000 deaths annually. Additionally, ninety-two percent of drowning deaths occur in low- and middle-income countries. Rapid detection and alert during drowning is critical to reduce the amount of drowning fatalities. Current drowning detection system approaches consist of two main categories, wearable systems and vision-based systems. Vision-based systems, using deep learning models along with cameras, are non-intrusive, comparatively inexpensive, don’t require as much maintenance, and are therefore the more effective approach for low-cost drowning detection as compared to wearable systems. This being said, current computer-vision models are too heavy and too slow for low-cost edge deployment due to high parameters requiring expensive hardware. To solve these current issues with vision-based systems, we created INO-YOLO. With the help of LGCBlock combined with an optimized Bi-directional Feature Pyramid Network for feature fusion, our model is able to effectively create a model with reduced computational cost, while maintaining similar accuracy and a low false-negative rate. To ensure the robustness of the model in various settings, the model was trained on an extremely diverse dataset consisting of images of various environments, lighting conditions, perspectives, subject ages, and devices. Despite being extremely lightweight, our model achieved exceptional performance. Its computational cost was lower than every other well-known published drowning detection model: 3 GFLOPS, 1.4 million parameters, and a file size of 2.95 MB, while maintaining a strong accuracy and recall of 0.9 and 0.86 respectively. Our model also allows for rapid real-time detection, with a high model FPS of 256, due to the reduced parameters. Due to this lightweightedness, the model is able to perform comfortably on inexpensive microcontrollers such as the Raspberry Pi Zero 2W, allowing for low-cost edge deployment. These results suggest the model’s strong potential for use in real-time drowning detection applications in low- and medium-income countries and its ability to generalize well to different environments, potentially contributing to the reduction of preventable drowning-related fatalities worldwide.