Navigation of robotic platforms in commercial horticulture: a comparative analysis of transformers for semantic segmentation

The article presents the results of research on the use of deep learning methods based on SegFormer models for semantic image segmentation and autonomous navigation of robotic platforms in rows of orchard plantings. The study compared different versions of SegFormer models pretrained on ADE20K and CityScapes. The data sets show a difference in accuracy of up to 4–7 % and diff er in the number of classes and the number of model parameters (from 3.7 million for B0 to 82 million for B5). For transfer learning of the models, a dataset was prepared and labeled. The dataset consisted of 1200 images of rows of orchard plantings with annotations for six classes of objects. These included the class Tree (apple trees with a height of ≥1.5 m), the class Near-Trunk (zones around the trunk, with a radius of 0.5 m), the class Pole (support structures with a height of ≥2 m), the class Sky (the sky area, including clouds), the class Track (row spacing, 3 m wide), and the class Background (the surrounding environment, other background objects) in Robofl ow. To expand the dataset and improve the quality of the models, we performed data augmentation (image rotation, brightness correction). The experiments showed that increasing the dimensionality of hidden layers in the SegFormer B0–B5 models enhanced feature extraction from images, correlating with the growth metrics for semantic segmentation. The comparative analysis of 12 versions of SegFormer models (B0–B5) identifi ed the optimal ratio between accuracy and performance. For instance, the B4 SegFormer version achieves the highest accuracy (Val Dice=0.7927) and is recommended for high-detail mapping tasks. The B0 SegFormer version ensures the highest processing speed (1.52 FPS) and is applicable for real-time navigation. The designed trajectory construction algorithm based on DBSCAN clustering and RANSAC approximation showed high efficiency in eliminating noise in segmentation and generating movement routes. Thus, robotic platforms are enabled to adapt to dynamic conditions, including changes in row geometry and background interference. The developed algorithm will improve positioning accuracy and reduce the dependence of robotic platforms on expensive multisensor systems.