Abstract: Severe challenges including difficult navigation path extraction, unsatisfactory real-time response and erroneous line switching during continuous field operation are brought about by the intricate and variable growth environment of honey melon plants across the entire growth cycle. To solve the above technical problems, a novel visual navigation path extraction approach for agricultural robots oriented to full-cycle growth characteristics of honey melon is proposed in this research. The improved LG-YOLOv8n-seg model is employed to accomplish accurate segmentation of inter-row regions in honey melon farmland. A multi-branch VanillaBlock named MB_VanillaBlock is constructed as the core backbone structure, and a shared lightweight segmentation head entitled SharedLite_Segment is specially designed. Both optimized modules are adopted to effectively elevate the overall reasoning efficiency of the segmentation network. Meanwhile, a boundary geometry enhancement module abbreviated as BGE_Module is integrated into the network framework, by which the segmentation precision of fuzzy and irregular boundary areas is remarkably strengthened. Valid experimental data are acquired to verify the comprehensive performance of the improved model. The parameter scale, computational load, mAP50 and BF_Score of LG-YOLOv8n-seg are measured as 1.437 M, 7.5 G, 96.10% and 61.17% respectively. In comparison with the original benchmark YOLOv8n-seg model with corresponding values of 3.258 M, 12 G, 97.65% and 57.68%, the total parameter quantity is decreased by 55.89%, and the computational consumption is lowered by 37.5%. The BF_Score indicator is raised by 3.49 percentage points, whereas merely a slight decline of 1.55 percentage points is generated in the mAP50 index. Valid feature points are screened and redundant noise points are discarded according to the obtained segmentation images. The least square algorithm is applied to fit smooth central paths of crop rows. The Dubins curve theory is introduced to generate stable line-switching trajectories. Steady and continuous path extraction tasks are completed for honey melon rows covering all different growth phases. The established LG-YOLOv8n-seg model is transplanted and deployed onto the NVIDIA Jetson Orin NX embedded development board. The actual running inference speed is tested up to 42.55 frames per second, and a performance improvement of 17.41% is achieved when compared with the original YOLOv8n-seg model. The average deflection angle of inter-row navigation paths throughout the whole growth period is controlled at only 0.82 degrees. Navigation deviations and the proportional values of deviations relative to actual inter-row width at each growth stage are all confined within reasonable permissible ranges. The practical working demands of visual navigation system in complex honey melon field scenarios can be fully fulfilled.