Answers are grounded in this paper and cite their sources.
• A lightweight ridge instance segmentation model (R-YOLOv11-Prune) was developed. • Large-kernel Ghost-DW and Dir-SE modules enhance ridge feature representation. • BN sparsity-guided pruning reduced parameters by 82% with 0.992 mAP maintained. • A dynamic centerline and piecewise fitting algorithm ensured smooth navigation lines. • The embedded system achieved 23.9 ms per frame with <1° heading angle error. Current rapeseed transplanting requires operators to walk alongside the machine to adjust its position and ensure planting accuracy. However, visual occlusion often makes it difficult for operators to judge the ridge direction, which reduces operation quality. Therefore, automatic ridge following is essential to reduce labor dependence and ensure operation quality, for which reliable visual navigation is the key. To meet this need, this paper proposes R-YOLOv11-Prune, a lightweight ridge instance segmentation model alongside a robust navigation line extraction algorithm, and validates the performance on the D Robotics RDK X5 edge platform. R-YOLOv11-Prune uses a large-kernel longitudinal Ghost-DW module and a direction-aware attention module to improve the feature representation of ridges. Furthermore, the model uses Batch Normalization sparsity-guided channel pruning to significantly reduce complexity while keeping accuracy. The proposed navigation line extraction algorithm combines dynamic centerline point extraction with piecewise curve fitting to generate smooth navigation lines, effectively reducing the impact of irregular boundaries and broken ridges. Experimental results demonstrate that R-YOLOv11-Prune achieves an mAP50 of 0.992, a Precision of 0.977, and a Recall of 0.993. With only 0.5 M parameters and 4.5 GFLOPs, the model reduces the parameter count and computational cost by approximately 82% and 56%, respectively, compared to the YOLOv11n baseline. Validation on the D Robotics RDK X5 platform shows that the model achieves an inference speed of approximately 23.9 ms per frame. Additionally, the navigation line extraction algorithm yields a mean heading angle error of 0.70°–0.87° and a mean pixel error of 3–4 pixels across diverse field scenarios. Compared with traditional Hough transform and RANSAC methods, the proposed algorithm reduces the heading angle error by over 49% and lateral pixel error by over 66%. These results prove the practical feasibility of the method for high-precision, embedded ridge following.
No takes yet. Share an insight, caveat, or question.
Han et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: