基于改进YOLO 11的烟苗垄作行导航线提取方法

Extraction Method of Navigation Lines for Tobacco Seedling Ridged Rows Based on Improved YOLO 11

摘要:

针对团棵期前烟苗在田间视觉导航中检测准确率低、模型参数量大等问题,本文提出一种基于改进YOLO 11的垄作物行导航线提取方法,有效压缩模型规模并提升田间移动平台作业导航精度。面向烟苗目标检测任务,设计NewBlock模块并引入LightAttention机制,分别强化YOLO 11主干网络特征提取能力与轻量化性能,提升检测效果。构建BiFPN_ADown融合结构替代Neck中PANet结构,实现参数压缩与特征保持平衡。采用Canny边缘检测算法提取烟苗外接轮廓计算其生理学质心,结合最小二乘法实现作物行导航线精确拟合。消融试验结果表明,改进模型精确率、召回率和平均精度分别提高3.9、0.8、1.9个百分点,计算量降低47.62%。作物垄上导航线拟合验证结果表明,平均角度误差与距离误差分别为0.91°与11.78像素。在复杂农田环境下,本文算法综合性能优异,可有效平衡检测速度与精度,为烟草及其他垄作物田间机器人视觉导航提供了新思路。

Abstract:

Aiming to address the problems of low detection accuracy and large model parameter size in field visual navigation for tobacco seedlings before the rosette stage,a ridge crop row navigation line extraction method was proposed based on an improved YOLO 11. This approach effectively compressed the model size and enhanced the navigation precision of field mobile platforms. To address the tobacco seedling target detection task,a NewBlock module was designed and a LightAttention mechanism was introduced to respectively enhance the feature extraction capability and lightweight performance of the YOLO 11 backbone network,thereby improving detection performance. The BiFPN_ADown fusion structure was constructed to replace the PANet in the Neck,achieving a balance between parameter compression and feature preservation. The Canny edge detection algorithm was employed to extract the external contours of tobacco seedlings and calculate their centroids. Subsequently,combined with the least squares method,precise fitting of the crop row navigation lines was achieved. The ablation experiment results demonstrated that the improved model achieved increases of 3.9,0.8,and 1.9 percentage points in precision,recall,and mean average precision,respectively,while reducing computational complexity by 47.62%. Validation results for crop ridge navigation line fitting indicated an average angle error of 0.91° and a distance error of 11.78 pixels. In complex farmland environments,the proposed algorithm exhibited superior comprehensive performance by effectively balancing detection speed and accuracy,providing a solution for the visual navigation of field robots in tobacco and other ridge crop operations.