Deep learning-based semantic segmentation with novel navigation line extraction for autonomous agricultural robots

In the rapidly evolving field of agriculture, the integration of autonomous systems is essential for enhancing productivity and efficiency. This study addresses the challenge of navigation line extraction in autonomous agricultural robots, a critical component for precise field operations. In this study, three deep learning-based semantic segmentation models, ENet, Deeplabv3+, and PSPNet, were trained on corn crop row images to extract the traversable path of the robot. Then, we proposed a novel navigation line extraction algorithm based on the Douglas–Peucker algorithm which uses the output mask from the semantic segmentation to extract the centre navigation line for robot guidance. The results of the experiments showed that PSPNet achieved the highest mean intersection over union (mIoU) of 96.50%, followed by Deeplabv3+ (96.30%) and ENet (95.13%). While ENet showed more consistent performance across various lighting conditions, the novel algorithm demonstrated remarkable accuracy, reducing angle errors to an average of 1.1°, 1.6°, and 1.6° for ENet, Deeplabv3+, and PSPNet, respectively, compared to 4.6°, 4.9°, and 6.7° using the baseline method. This improvement is critical for ensuring precise and stable robot navigation in agricultural fields. Beyond its technical contributions, this study offers practical implications for real-world deployment, ensuring reliable operation in dynamic agricultural environments. By addressing limitations in conventional navigation techniques, the proposed method enhances robot maneuverability in corn crop fields, particularly under challenging lighting conditions. However, its application across different crop types and terrains, further optimizing real-time processing efficiency, are yet to be explored. This work represents a significant step toward achieving robust, autonomous navigation in precision agriculture.