Authors

Abstract

The seedling stage is a relatively short but critical period in the maize growth cycle, during which rapid weeding, pesticide application, and fertilization are required. Crop row centreline extraction technology based on computer vision can assist agricultural machinery in performing automated field operations, thereby improving work efficiency. Therefore, improving the accuracy of crop row centreline extraction imposes higher requirements on existing models and algorithms. In this study, a model named Synergistic Attention YOLO (SA-YOLO) was proposed. The feature extraction capability for crop rows was enhanced by introducing the spatial and channel synergistic attention module and designing the A2C2f module integrated with switchable atrous convolution. In addition, a crop topological skeleton feature point extraction strategy and a collaborative fitting method combining principal component analysis and least squares was integrated to accurately extract crop row centrelines from the skeleton feature points of crop pixels within the segmented region of interest. Comparative experiments showed that the SA-YOLO model achieved competitive performance, with an average precision at IoU threshold 0.5 (AP50) of 99.1%, and a mask Intersection over Union (IoUmask) of 83.8%, while maintaining a lightweight architecture with only 3.72 M parameters. The mean angle error (A_mean) of crop row centerline extraction was 0.590°, and the mean normalized lateral error (L_mean^norm) was 0.357%. This study provides a new approach for exploring crop row centerline extraction in agriculture and can further enrich the theoretical and technical foundation for visual navigation of agricultural robots.

Abstract in Chinese

苗期是玉米生长周期中相对较短但关键的时期，需要快速完成除草、施药、施肥等工作。基于计算机视觉的作物行中心线提取技术可以辅助农业机械进行自动化田间作业提高工作效率。因此，作物行中心线提取精度的提升对现有模型与算法提出了更高的要求。本研究提出了一种名为SA-YOLO的模型，通过引入SCSA注意力模块和设计的A2C2f_SAConv模块增强了对作物行特征的提取能力。此外，还集成了作物拓扑骨架特征点提取策略与PCA-LSM协同拟合方法，通过分割输出的ROI区域内作物像素的骨架特征点来精确提取作物行中心线。对比实验表明，SA-YOLO模型实现了具有竞争力的性能，其AP50为99.1％，IoUmask为83.8％，同时保持轻量级架构，参数仅为3.72M。作物行中心线提取产生的平均角度误差为0.590°，归一化横向误差为0.357%。该研究为探索农业作物行中心线提取提供了新的思路，可以进一步丰富农业机器人视觉导航的理论和技术基础。