Binocular stereo vision-based relative positioning algorithm for drone swarm.

Area of Science:

• Robotics
• Computer Vision
• Aerospace Engineering

Background:

• Binocular vision-based Unmanned Aerial Vehicle (UAV) formation flight faces challenges with high computational complexity and poor real-time performance.
• Existing methods often struggle to balance accuracy and efficiency for dynamic UAV operations.

Purpose of the Study:

• To develop an efficient and precise UAV localization algorithm for formation flight.
• To improve the real-time performance of binocular vision systems in UAVs.

Main Methods:

• Optimization of the YOLOv5s model using lightweight design principles, creating Yolo-SGN.
• Utilizing Yolo-SGN for targeted feature point extraction in binocular images, combined with the Oriented FAST and Rotated BRIEF (ORB) algorithm.
• Implementing a binocular vision localization model using extracted feature points to compute 3D coordinates for UAV positioning.

Main Results:

• The Yolo-SGN model achieved a 65.5% reduction in parameters, a 62.7% reduction in FLOPs, and a 1.8% increase in accuracy compared to the original YOLOv5s.
• Feature point matching computations were reduced to one-quarter of the original ORB algorithm's load.
• The proposed algorithm demonstrated exceptional precision and real-time capabilities in UAV localization.

Conclusions:

• The lightweight Yolo-SGN model significantly enhances the efficiency of object detection for UAV applications.
• The integration of Yolo-SGN with ORB and binocular vision localization provides a robust solution for real-time UAV formation flight.
• This approach offers a promising direction for improving the performance and applicability of autonomous UAV systems.