Smart vineyard row navigation: A machine vision approach leveraging YOLOv8

• Machine vision-based system effectively navigates in GNSS-limited environments. • Custom dataset supports training for precise vine end detection. • Vine end detection with YOLOv8 achieves high precision under diverse conditions. • Innovative lateral and heading error metrics developed for vineyard navigation. • Precise row end detection method developed for safe EV stopping. This study presents a novel machine vision-based framework for autonomous navigation inside vineyard rows, addressing the challenges posed by unreliable global navigation satellite system (GNSS) signals in such environments. By focusing on the detection of vine ends, a critical dataset was compiled from images captured in a vineyard. This dataset served as the foundation for training four YOLOv8 object detection models, with the best-performing model being YOLOv8m-vine-classes. The YOLOv8m-vine-classes model achieved a precision of 95 % and an mAP50 of 93.7 %, demonstrating its effectiveness in detecting vine ends for accurate autonomous navigation. The model enabled the development of an algorithm capable of generating lateral and heading error information—key metrics for precise and safe autonomous navigation. Real-time field experiments underscore the system’s effectiveness, with the root mean square error (RMSE) for lateral error consistently near the 5 cm mark and that for heading error within 2 degrees. These findings indicate the system’s adeptness at maintaining both proximity to the desired path and correct orientation throughout the navigation process. Additionally, the system demonstrated robust performance in detecting the ends of tree rows and calculating the distance to the final end post. Tests showed that the ’End of path’ flag is generated as soon as the vine ends exit the image frame, with initial distance measurements falling within the 2.5–3 m range. This ensures that the EV can safely stop within the next 2–3 s at a distance of approximately 1 m from the end of the row. The maximum error observed for distance measurement during these tests was 0.27 m, validating the system’s accuracy and reliability.