Extracting the central crop row with CCRDNet for universal in-row navigation in agriculture

Deep learning has recently shown strong potential in crop row detection for navigation line extraction. However, existing approaches often rely on dataset-specific customization and extensive image preprocessing, limiting their practicality in real-world agricultural scenarios. In contrast, human operators can instinctively navigate machinery by simply following the central crop row. Inspired by this observation, we propose a novel strategy that directly extracts the central crop row as the navigation line. To support this paradigm, we introduce a three-class annotation scheme-background, vegetation, and central crop row-where the vegetation class serves as an auxiliary supervisory signal to provide structural constraints and guide accurate localization. A consistent annotation width of crop row is applied across all samples to enable the model to learn invariant structural features. We develop CCRDNet (Central Crop Row Detection Network), which predicts the central row position and subsequently fits the navigation line using the least-squares method. A dataset of 7,367 images comprising eight crop types across diverse environments was collected, yet only 400 images-from two crop types in eight environments-were used for training. Despite the limited supervision, the proposed method achieved a navigation line extraction accuracy of 95.57% with an average angle error of 1.13°. CCRDNet is lightweight, requiring only 0.033M parameters, and operates at 86.76 FPS on an RTX 3060 GPU and 48.78 FPS on a Jetson Orin NX. These results demonstrate that the proposed approach not only simplifies the navigation pipeline but also enables zero-shot generalization across previously unseen environments, fully satisfying the real-time requirements of agricultural machinery.