Progetto di un sistema di visione artificiale su piattaforma embedded per veicoli da competizione Formula Student Driverless

Formula Student is an international motorsport competition in which a team of students design and build a single-seater racing vehicle. Starting in 2017, the driverless category has fostered the development of self-driving cars racing on a track bounded by cones of known color and size. This thesis proposes a low-cost, low-complexity perception system for recognizing those cones and determining their coordinates in the vehicle’s reference frame. The system employs a monocular camera and an NVIDIA Jetson Orin Nano Developer Kit, which together cost around €1000. The algorithm runs YOLO object detection on images from a calibrated camera and applies prior knowledge of cone geometry to compute positions using the pinhole camera model. The system achieves an mAP@50:95 of 46.73% with YOLOv10n and 48.53% with YOLOv10s in INT8 precision. The localization accuracy is evaluated with a root mean square error of 0.840 m for the nano model and 0.760 m for the small model. With a power consumption of approximately 7 W, the system delivers frame rates of 31.80 fps and 29.90 fps, respectively. Such results suggest the system can support real-time, safe autonomous navigation in the Formula Student Driverless scenario.