Revolutionizing Safety: How AI-Powered Technology Detects Dangerous Rip Currents Instantly

Article

Object Detection and Image Classification

January 28, 2025

To help prevent injuries from rip currents, the Korea Hydrographic and Oceanographic Agency (KHOA) has started using object detection and image classification technology in real-time monitoring systems. This approach uses AI to quickly spot rip currents in images and alert safety personnel, allowing for speedy action.

Rip currents are strong water channels that flow from the beach to the open sea. They can be deadly, especially in summer when many people swim. These currents often appear suddenly, even in calm weather, making them hard to see. To tackle this problem, KHOA implemented advanced AI and computer vision technologies in high-risk areas, enhancing the ability to detect these currents in real time.

Automatic Rip Current Detection Technology

KHOA developed this detection technology at Haeundae Beach, known for its frequent rip currents and past accidents. They used Ultralytics’ YOLOv8 (You Only Look Once, version 8), an advanced object detection model that processes entire images quickly, unlike older models that scan pictures multiple times. This speeds up detection, making it suitable for real-time use.

The YOLOv8 model works by dividing images into grids and predicting both object presence and location. Here’s how it operates:

• **Image Segmentation**: The image is split into grids that check for object presence.
• **Bounding Box Prediction**: Each grid suggests boxes surrounding potential objects.
• **Class Probability Prediction**: It assesses the likelihood that each box holds a specific object.
• **Non-Maximum Suppression (NMS)**: It keeps only the box with the highest likelihood, discarding the rest.

To train the YOLOv8 model, researchers used three years of images from four CCTV cameras at Haeundae Beach. They classed the images as either “RipCurrent,” where a current was clear, or “RipDoubt,” where it was less obvious. They also included images without rip currents as controls.

Learning, Verification, and Testing Images

The final dataset included 58,000 images, which were split into training, verification, and testing groups. The model was trained with the learning data and validated using the test set. Its performance was measured by checking precision, recall, and mean average precision (mAP) across multiple categories.

The results showed impressive scores: precision at 90.3%, recall at 91.7%, and mAP at 95.0%. These figures indicate excellent accuracy in detecting rip currents.

Conclusion

Rip currents pose a significant risk at beaches, often leading to tragic accidents. Quick and accurate detection technology is crucial for safety. KHOA’s automatic detection system, paired with real-time monitoring, enhances responses to dangerous currents. This technology holds promise for use across other beaches, aiming to reduce fatalities and improve ocean safety.