LCD-Net: An End-to-End Agricultural Robot Navigation Line Extraction and Crop Row Detection Neural Networks

The autonomous navigation of agricultural robots relies on precise, real-time visual perception. However, traditional methods are computationally expensive and sensitive to complex environmental changes such as lighting and weather, making it difficult to meet the demands of practical applications. To address these limitations, this paper proposes an end-to-end visual navigation approach based on Liquid Neural Networks (LNNs). Given a raw image input, the model directly outputs both the navigation path and crop row regions, significantly simplifying the navigation pipeline. The model employs a “left-center-right” semantic perception structure, generating three outputs: left crop row, right crop row, and central navigation path. This structure enhances robustness and adaptability under challenging conditions such as occlusion, curves, and dense crops. On the CRDLD agricultural navigation dataset, the proposed LCD-Net achieves a leading mIoU of 34.41 %, surpassing classical architectures by 14 percentage points. Regarding practical deployment, the model exhibits outstanding computational efficiency with only 6.49M parameters, representing a reduction of approximately 75 % compared to mainstream methods, while maintaining a peak GPU memory footprint of merely 158.74 MiB. These experimental results validate LCD-Net as a precise, efficient, and robust autonomous navigation solution that demonstrates superior performance and generalization on platforms with limited resources.