A navigation line extraction method based on semantic segmentation and centerline fitting

This paper aims to overcome the limitations of accuracy deficiency and real-time performance deficits in traditional farmland navigation line extraction under complex scenarios (e.g., sudden illumination variations, crop row fragmentation, and weed coverage), this study introduces an integrated navigation line extraction approach combining enhanced PSPNet semantic segmentation and centerline fitting. Firstly, to satisfy stringent real-time constraints for navigation line extraction, the backbone network was replaced with MobileNetV1, thereby achieving model lightweighting. Secondly, to enhance segmentation precision for crop row edges in challenging environments, a hybrid boundary loss function was incorporated, resulting in smoother and more accurate segmentation boundaries that adhere more closely to the actual crop row contours. Subsequently, the PSPNet-generated crop rows were binarized and integrated with a least squares method to derive centerlines for navigation path generation. Experimental results demonstrated that the refined PSPNet method achieved a mean accuracy of 93.16% with an average processing time of 80 milliseconds per image, significantly improving the real-time capability of navigation line extraction compared to conventional visual algorithms. These outcomes confirm that the proposed approach meets field operation accuracy requirements while delivering high-precision path reference benchmarks for agricultural machinery navigation.