马驰,董子扬,陈志军,朱悦,石复习.基于根点替代的猕猴桃果园行间导航线生成方法研究[J].干旱地区农业研究,2021,39(5):222~230
基于根点替代的猕猴桃果园行间导航线生成方法研究
Research on navigation line generation of kiwi orchard between rows based on root point substitution
  
DOI:10.7606/j.issn.1000-7601.2021.05.29
中文关键词:  猕猴桃果园  导航线生成  树干识别  根点替代
英文关键词:kiwi orchard  navigation line generation  trunk detection  root point substitution
基金项目:陕西省重点产业创新链(群)项目(2020ZDLNY07-05)
作者单位
马驰 College of Mechanical and Electronic Engineering,Northwest A&F University, Yangling, Shaanxi 712100, China 
董子扬 College of Mechanical and Electronic Engineering,Northwest A&F University, Yangling, Shaanxi 712100, China 
陈志军 College of Mechanical and Electronic Engineering,Northwest A&F University, Yangling, Shaanxi 712100, China 
朱悦 College of Mechanical and Electronic Engineering,Northwest A&F University, Yangling, Shaanxi 712100, China 
石复习 College of Mechanical and Electronic Engineering,Northwest A&F University, Yangling, Shaanxi 712100, China 
摘要点击次数: 282
全文下载次数: 252
中文摘要:
      针对棚架式猕猴桃果园树干分立、树干与树冠分离的生理特点,提出基于深度学习的目标检测框底边中点替代果树根点的导航特征目标检测方法。使用LabelImg标注预处理图像中的树干信息,生成猕猴桃树干数据集,基于Faster R-CNN构建目标检测模型,实现对猕猴桃果园行间有效范围的树干识别,利用根点替代方法确定树干根点的定位基点坐标,基于三次样条插值法提取树行线,根据最小二乘法拟合生成导航线。试验结果表明,生成的目标检测模型在地膜、杂草和土壤等果园行间环境下,树干正确识别率分别为90.6%、90.1%、89.4%;树干像素高度大于300像素、200~300像素、100~200像素的正确识别率分别为94.6%、89.9%、85.6%,表明选择树干作为果园行间视觉导航标识,有效利用了深度学习的稳定性,提高了导航目标识别的准确率;根点替代简化了导航定位基点生成算法,获取的果树定位基点与实际树干根点的平均横向像素误差占比为1.3%,平均纵向像素误差占比为2.2%,平均实际距离误差为0.102 m,导航定位基点精度较高;生成的导航线平均横向像素偏差为5.1像素,实际平均横向偏差为0.052 m,获取的果园行间导航路径满足行间导航要求。
英文摘要:
      In orchard visual navigation, the selection and extraction of navigation signs are the key issues in the generation of navigation lines, which determine the robust and accuracy of the generation of navigation lines in the orchard. Based on the physiological characteristics of the separation of the trunk as well as the separation of the trunk and the crown of the trellis kiwi fruit orchard,and using the deep learning method of the bottom mid\|point of the target detection frame, this study proposed a navigation mark generation method to replace the root point of the fruit tree. LabelImg was used to annotate the trunk in the preprocessed image and generate a kiwi trunk data set. Object detection model based on Faster R-CNN was built to realize the trunk recognition of the effective range of kiwi fruit orchard between rows. The root point substitution method was employed to determine positioning base point of the trunk root point. The base point coordinates were extracted from the tree line based on the cubic spline interpolation method, and the navigation line was generated according to the least squares method. The test results showed that in the inter\|row environment of mulch plastic film, weeds and soil, the correct recognition rates of trunks were 90.6%, 90.1%, and 89.4%. The trunk pixel height in the image was greater than 300 pixels,trunk pixel height was between 200~300 pixels, and trunk pixel heights between 100~200 pixels, the correct recognition rates of small targets trunk were 94.6%, 89.9%,and 85.6%. Selecting trunks as the orchard visual navigation target showed effective use of the stability of the depth of learning and improved accuracy of navigation target recognition. Root point substitution simplified the navigation positioning base point generation algorithm. The average horizontal pixel error between the fruit tree positioning base point and the actual trunk root point accounted for 1.3%, average vertical pixel error accounted for 2.2%, and the actual average distance error was 0.102 m. The positioning base point of navigation had high accuracy. The average horizontal pixel deviation of the generated inter\|row navigation line was 5.1 pixels, and the actual average horizontal deviation was 0.052 m. Navigation path generation of kiwi orchard between rows met the requirements.
查看全文  查看/发表评论  下载PDF阅读器