Anchor-Based In-Row navigation line detection under diverse agricultural conditions

Accurate in-row navigation is fundamental for the autonomous operation of agricultural machinery. However, real-world fields pose significant challenges, such as variable lighting, discontinuous rows, and diverse growth stages. While deep learning shows promise, existing models often struggle to generalize across unseen environments. To address these limitations, this study proposes AgriLaneNet, a streamlined anchor-based framework for direct navigation line detection in agricultural in-row navigation. Instead of explicitly segmenting all crop rows or reconstructing the complete crop-row geometry, AgriLaneNet reformulates the task as the estimation of ordered anchor points along predefined horizontal anchors. These points are then fitted by least-squares regression to obtain the frame-wise local navigation line for row-following guidance. This design simplifies the navigation pipeline by reducing dependence on crop type identification or multi-row detection, and complex geometric post-processing. Furthermore, a Discrete Flow Mapper (DFM) module is introduced, which reformulates coordinate regression as discrete probability distribution estimation to improve prediction precision and robustness. Trained on only 800 images from 16 distinct scenes, AgriLaneNet achieves 97.53% accuracy on a diverse test set of 8,145 images, with a mean angular error of 0.94°. The model also demonstrates strong cross-domain generalization, achieving 96.08% accuracy on unseen domains, and high computational efficiency, reaching 171.20 FPS on GPU. In addition, video-based evaluations further indicate that the proposed method can maintain stable navigation line detection under continuous field observations. These results demonstrate the robustness, generalization ability, and practical potential of AgriLaneNet for real-time agricultural navigation in complex field conditions.