Enhancing UAV Visual Landing Recognition with YOLO's Object Detection by Onboard Edge Computing.

Area of Science:

• Robotics and Computer Vision
• Aerospace Engineering

Background:

• Unmanned Aerial Vehicles (UAVs) require efficient onboard object detection for navigation and reconnaissance.
• Current systems face challenges in real-time processing, data transmission, and accuracy in diverse environments.

Purpose of the Study:

• To enhance UAV visual capabilities for landing and reconnaissance missions.
• To improve object detection speed and accuracy using edge computing.
• To reduce data transmission and processing time for ground stations.

Main Methods:

• Implementing You Only Look Once (YOLO)-based object detection with TensorRT acceleration on an onboard edge computer.
• Utilizing an automated visual tracking gimbal camera control system.
• Employing multithread programming for efficient image transmission.
• Comparing four YOLO models and applying YOLOv4-tiny to a real-world field test.

Main Results:

• Achieved high frames per second (FPS) rates with YOLO accelerated by TensorRT on UAVs.
• Demonstrated satisfactory mean average precision (mAP) with lightweight edge computing.
• Successfully applied trained YOLOv4-tiny models to recognize landing spots over 100 km away in unknown environments.
• Confirmed the feasibility of using NVIDIA Jetson Xavier NX with YOLO achieving over 35 FPS.

Conclusions:

• The proposed approach significantly enhances UAV real-time object detection and visual capabilities.
• The system demonstrates successful autonomous landing and reconnaissance potential in new environments.
• Integration of YOLO, TensorRT, and edge computing provides a viable solution for advanced UAV missions.