A real-time orchard navigation path extraction method using semantic segmentation and pixel scanning

Autonomous navigation technology plays a pivotal role in facilitating intelligent operations within orchard environments. To address the challenges of real-time performance and generalization, this study proposed a novel real-time visual navigation path extraction method that integrated semantic segmentation with pixel scanning techniques. The research introduced the MCFF-Segformer semantic segmentation model, which was trained and validated on a comprehensive orchard road dataset. Building upon the segmentation results, the navigation paths were extracted through an innovative approach combining pixel scanning technique with cubic spline interpolation. Experimental evaluations demonstrated the effectiveness of our approach, with the MCFF-Segformer model achieving remarkable performance: an MIoU of 88.97%, an MPA of 94.91%, and a processing speed of 30.4 FPS, thereby satisfying both accuracy and real-time operational requirements. When implementing a scanning interval of 90 pixels, the system maintained an average pixel error of merely 12.2 pixels, corresponding to a physical distance error of 0.04 m. These results indicated that the proposed strategy significantly ensured the accuracy of path extraction, thereby contributing to enhanced navigation precision in orchard environments.