Visual navigation path extraction method for hami melon fields at all growth stages using LG-YOLOv8n-seg

Honey melon cultivation has posed severe challenges to autonomous agricultural robots, due to the variable environment of plant growth over the entire growth cycle. Plant leaves continuously cover the inter-row soil areas, as the vines grow and spread progressively in different periods, thus blurring the dividing lines between crop rows. Dynamic canopy closure can also aggravate three engineering troubles: difficult extraction of navigation path, unsatisfactory real-time response, and erroneous line switching during operation in the field. In this study, the visual navigation path extraction was proposed for the full-cycle growth of honey melon in agricultural robots. An improved lightweight instance segmentation model, termed Lightweight Geometry-enhanced YOLOv8n-seg (LG-YOLOv8n-seg), was also employed for the accurate pixel-level segmentation of inter-row regions in honey melon farmland. A multi-branch VanillaBlock, named MB_VanillaBlock, was constructed as the backbone structure to extract multi-scale spatial features of crop and soil areas. Meanwhile, a shared lightweight segmentation head with SharedLite_Segment was specially designed to avoid repeated feature calculation. Both optimal modules were adopted to effectively elevate the overall inference efficiency of the segmentation network and baseline accuracy. Furthermore, a boundary geometry enhancement module, abbreviated as BGE_Module, was integrated into the network framework. Geometric constraints were imposed on ambiguous boundary predictions. This module remarkably strengthened the segmentation precision of fuzzy and irregular boundary areas blocked by dense melon leaves. A series of valid experimental data was acquired to systematically verify the performance of the improved model against the original YOLOv8n-seg benchmark. The parameter scale, computational load, mean average precision under 0.5 intersection over union threshold (mAP50), and boundary fitting score (BF_Score) of LG-YOLOv8n-seg were measured as 1.437 M, 7.5 G, 96.10%, and 61.17%, respectively. In YOLOv8n-seg, those were the 3.258 M parameters, 12 G computational load, 97.65% mAP50, and 57.68% F₁ Score, while the total parameters decreased by 55.89% and the computational consumption was lowered by 37.5%, compared with the original YOLOv8n-seg model. The BF_Score indicator was raised by 3.49 percentage points, whereas a slight decline of 1.55 percentage points was generated in the mAP50 index. The trade-off demonstrated the effective balance between lightweight and segmentation boundary accuracy. Valid feature points were then optimized from the foreground–background separation using the segmentation images. Redundant noise points were removed against weeds and uneven ground. The least squares algorithm was applied to fit smooth central navigation paths of crop rows using the points retained from the effective boundary feature. Furthermore, Dubins' curve theory was introduced to generate smooth and stable line-switching trajectories without sharp steering angles. Dubins curve respected the minimum turning radius of the agricultural vehicle to produce kinematically executable paths, thereby directly overcoming the erroneous line switching. Steady and continuous path extraction tasks were consequently completed for honey melon rows at different growth phases, from the seedling to the fruit ripening stage. The LG-YOLOv8n-seg model was established and then deployed onto an NVIDIA Jetson Orin NX embedded board in field real-time tests. The actual running inference speed was evaluated up to 42.55 frames per second, achieving a 17.41% performance, compared with the original YOLOv8n-seg model under the same embedded hardware environment. The average deflection angle of inter-row navigation paths was controlled at only 0.82° in the whole growth period. Permissible ranges of agricultural vehicles were obtained for the navigation deviations and their proportional values relative to the actual inter-row width at each growth stage. The findings can fully meet the practical demands of a visual navigation system under complex honey melon field scenarios with a dynamically growing plant canopy.