Exploring the Impact of Keypoint Precision on Vision-Based Autonomous Navigation
João Aires Marsicano, Felipe Tourinho, Marco S. Tayar, Marcelo Becker · 2025
Autonomous aerial navigation relies heavily on accurate real-time pose estimation. In this work, we investigate how the precision of keypoint detection impacts the performance of vision-based autonomous navigation systems. Using YOLO Pose v11 and a Perspective-n-Point (PnP) solver, we evaluate whether bounding box corners or keypoints yield more reliable distance estimations for UAV landings in dynamic environments. The study is framed within the high-demand context of the RoboCup Brasil Flying Robot Trial League and demonstrates the deployment of a full onboard solution powered by a Jetson Orin Nano. Through controlled experiments and a carefully annotated dataset, we show that keypoint-based predictions outperform bounding box estimates, especially under challenging angles and distances. Our findings provide critical insights for developers seeking to enhance real-time navigation accuracy in embedded robotics, offering a reproducible methodology with measurable gains in performance.