Perspective-Robust Square Gate Detection Using YOLO for Autonomous Drone Racing

Abstract

Gate detection using deep learning methods such as YOLO has become popular in recent years due to aerial robotics competitions, including AlphaPilot, IROS Autonomous Drone Racing, Game of Drones, and the A2RL × DCL. In these competitions, unmanned aerial vehicles (UAVs) have to navigate at high speed through a sequence of gates, where accurate detection is essential for navigation planning and reliable pose estimation. However, conventional YOLO detectors rely on bounding boxes, which become inadequate under perspective variations, such as rotations, oblique viewpoints, or scale changes, leading to poor agreement with the actual gate geometry. To tackle this limitation, we propose a novel perspective-aware approach that employs geometric constraints extracted from the four corners obtained via dual bounding-box detections. The method uses bounding-box detections to reconstruct a polygon representation of the gate, enabling a more accurate description of its shape across varying viewpoints.
The results show that our proposed method achieves a mean IoU score of 0.95 across three scenarios compared with traditional YOLO-based methods. Furthermore, our approach sustains real-time performance at 150 fps, keeping it suitable for implementation into autonomous drone racing systems.