This article is devoted to the development of a method for autonomous navigation of robotic platforms for industrial horticulture using semantic segmentation based on the SegFormer architecture. For training the models, a dataset including 1200 RGB images of orchard rows was used. The data were augmented (rotation ± 15°, brightness correction ± 15%) to enhance the model's robustness to variations in shooting conditions. Annotation was carried out by marking six classes of objects, including «Track» (motion trajectory) and «Tree» (trees), with class distribution balanced (70/15/15 for training/validation/testing). A comparative analysis of six model variants (B0-B5) was conducted, which revealed the optimal balance between accuracy (SegFormer-B5: Val mIoU = 0.59) and speed (SegFormer-B0: 1.52 FPS). The employed image processing methods (median filtering, spline approximation) ensured the smoothing of the motion trajectory. The proposed approach enables real-time trajectory planning and obstacle avoidance, ensuring safe and efficient navigation in dynamic orchard environments. Practical recommendations based on the study results include the use of SegFormer-B0/B1 for real-time navigation and SegFormer-B4/B5 for mapping tasks. The results of the work confirm the potential of applying SegFormer models in agricultural robotics for autonomous navigation in orchard rows.