Search and Rescue (SAR) missions in open water frequently face challenges such as low light, water surface reflection, obstruction, wave interference, and complex backgrounds, making it difficult for rescue personnel to spot drowning victims in a timely manner. To improve water rescue efficiency and real-time recognition capabilities, our team has developed the "Multi-modal AI-assisted UAV System for Real-time Drowning Detection". This system combines Unmanned Aerial Vehicle (UAV) aerial patrol, multi-sensor data integration, and AI image recognition technology to assist rescue units in quickly searching for suspected drowning targets in open waters such as seas and rivers.

Based on a multi-rotor UAV platform under 15 kilograms, this system integrates RGB Imaging, Thermal Imaging, Distance Sensing, and Geolocation capabilities. By utilizing Edge Computing and real-time image recognition models, it reduces reliance on cloud computing and ground equipment, thereby enhancing on-site deployment flexibility and response speed.
  Multi-Modal Sensor Integration

The system integrates RGB imaging, thermal imaging, distance sensing, and geolocation information to adapt to different water environments and lighting conditions. This enhances target search and interpretation capabilities, assisting rescue personnel in pinpointing the location of suspected drowning victims.

System Integration and Real-Time Communication

Through a central control module, this system integrates sensor data, mission control, flight behavior, and communication transmission, coordinating data exchange between various modules on the UAV and the ground control station. Detection results and mission status can be transmitted back to the operational interface in real-time, allowing rescue personnel on-site to quickly grasp target locations, image information, and mission progress, thereby improving the real-time response capability of water rescues.

Edge Computing and Data Fusion Architecture

To reduce reliance on cloud computing and remote servers, the system adopts an edge computing architecture, integrating multi-modal sensing, image preprocessing, AI recognition, and real-time data transmission onto the UAV platform. Through Data Fusion of RGB imaging, thermal imaging, distance measurement, and positioning information, detection stability and environmental adaptability are enhanced. Even in environments with restricted communication or low bandwidth, it can maintain real-time recognition and mission execution efficiency.

AI Model Training and Dataset Construction

This study utilized multi-rotor UAVs to collect images of real personnel simulating drowning scenarios, covering both river and marine environments, and incorporating various flight altitudes, shooting angles, body postures, and clothing conditions. After image cropping, preprocessing, and manual annotation, a total of 7,103 training images were established. The YOLOv8n Object Detection Model was employed for training to improve recognition capabilities for small targets, low contrast, and complex water surface backgrounds.

Real-Time AI Image Recognition

The system incorporates a deep learning image recognition model for the real-time detection of suspected drowning targets in open waters such as rivers and coasts. Through AI-assisted interpretation, the burden of manual visual searching is reduced, helping rescue personnel discover targets faster within the golden rescue time.

Stabilization and Image Preprocessing

To enhance image quality and recognition reliability, the system combines a stabilization module with Image Preprocessing workflows to reduce the impact of UAV flight shaking, water surface reflections, image blur, and noise on recognition results. This design helps improve the stability of real-time detection in water environments, making it particularly suitable for low-altitude flights, hovering, and dynamic SAR missions.

High-Accuracy Recognition Performance

After AI model training and testing, the system achieved a 98% detection accuracy rate, an mAP@0.5 of 0.991, and a peak F1-score of 0.97, demonstrating excellent real-time recognition performance and application potential.

Enhanced Adaptability to Complex Environments

When facing low light, water surface reflections, obstructions, complex backgrounds, and low-contrast targets such as dark clothing, the system can still maintain stable detection performance, contributing to the enhanced reliability and practicality of open water SAR missions.

Application Benefits

The Multi-modal AI-assisted UAV System for Real-time Drowning Detection can be applied to water rescue missions in seas, rivers, lakes, reservoirs, and flood-prone areas. Through rapid UAV deployment and aerial perspectives, it can assist rescue units in expanding the search range, shortening target discovery time, and improving rescue decision-making efficiency.

This technology can serve as a vital auxiliary tool for fire departments, coast guards, rescue teams, and local governments in water safety management, helping to strengthen real-time monitoring, disaster response, and water accident handling capabilities.
 

 

 

Technological Value

Continuous investment in the integration of AI, UAVs, and water SAR technologies drives smart, real-time, and highly efficient public safety applications. Through multi-modal sensing, AI image recognition, and real-world field verification, this technology not only elevates the application value of UAVs in water rescue missions but also establishes a more resilient solution for future smart disaster relief and public safety technologies.

International Journal Recognition

The research results related to this technology were published in May in the SCIE Q1 international journal Drones, under the title “Research on Real-Time Drowning Detection in Open Water Using Unmanned Aerial Vehicles and Artificial Intelligence Image Recognition”. The study focuses on the application of UAVs combined with artificial intelligence image recognition for real-time drowning detection in open waters, validating the system's performance through actual river and sea scenarios. This achievement demonstrates our technical team's R&D capabilities in AI UAVs, water SAR, and public safety technology applications, and also represents the technology's development potential for further promotion into actual rescue fields and international application contexts.