| This study evaluates YOLO (You Only Look Once) based object detection for autonomous Formula Student (FS) racing, focusing on the Leiria Academic Racing Team (LART) requirements. Using the Formula Student Objects in Context (FSOCO) dataset, we benchmarked successive iterations of the YOLO architecture, specifically versions v8, v9, and 11, across a resolution gradient ranging from 640 to 1280 px. While hyperparameter tuning yielded marginal gains, the most significant performance breakthrough resulted from a targeted data refinement algorithm that identifies and discards annotations for excessively small cones. This pre-processing step mitigated noise from distant, low resolution objects, substantially improving model precision and stability. Results show that optimized YOLOv8s and YOLO11s models provide the best balance of accuracy and real-time inference speed, meeting the rigorous latency constraints of autonomous track navigation. Our method achieved a score above 90\% mAP50 with a frame rate of 150fps, exceeding the requirements requested by LART, and it has been integrated into the LART autonomous driving prototype's perception software, having been successfully validated during the team's participation in an international Formula Student competition. |
*** Title, author list and abstract as submitted during Camera-Ready version delivery. Small changes that may have occurred during processing by Springer may not appear in this window.