基于特征点邻域Hough变换的水稻秧苗行检测  

Detection of Rice Seedling Rows Based on Hough Transform of Feature Point Neighborhood

作　　者：王姗姗 余山山 张文毅 王兴松[1] WANG Shanshan;YU Shanshan;ZHANG Wenyi;WANG Xingsong(School of Mechanical Engineering,Southeast University,Nanjing 211189,China;Nanjing Institute of Agricultural Mechanization,Ministry of Agriculture and Rural Affairs,Nanjing 210014,China)

机构地区：[1]东南大学机械工程学院,南京211189 [2]农业农村部南京农业机械化研究所,南京210014

出　　处：《农业机械学报》2020年第10期18-25,共8页Transactions of the Chinese Society for Agricultural Machinery

基　　金：国家重点研发计划项目(2016YFD020060502);中国农业科学院基本科研业务费项目(Y2019XK11)。

摘　　要：水稻秧苗行检测对于精准农业和自动导航至关重要,为此提出一种基于特征点邻域Hough变换的水稻秧苗行检测方法,该方法可以有效解决杂草密度分布、光照强度和秧苗行曲率变化等因素对秧苗行检测的影响。该方法主要包括3个步骤:水稻秧苗行图像数据库的建立、水稻秧苗特征点提取和秧苗行中心线识别。首先,在杂草萌发期建立水稻秧苗在不同光照条件(晴、阴天)、不同杂草密度分布和不同秧苗生长状况的水稻秧苗行图像数据库;然后,采用基于Faster RCNN网络的秧苗检测模型获得水稻秧苗的特征点,即预测结果的中心点;最后,采用提出的基于特征点邻域的Hough变换算法识别秧苗行中心线。实验表明,本文方法对测试集秧苗行平均识别准确率达到92%,对不同杂草密度分布的秧苗行平均识别精度小于0.5°,对孤立的杂草噪声和光照变化不敏感,对曲率较大的秧苗行也能准确识别,具有较好的鲁棒性和识别精度。The detection of rice seedling rows is essential for precision agriculture and automatic navigation.A method based on Hough transform of feature point neighborhood was proposed to detect rice seedling rows,which can effectively solve the effects of weed distribution with different densities,different light intensities,curvature changes of seedling rows and other factors.The method had three main steps:the establishment of images database of rice seedling rows,feature point extraction of rice seedlings and the recognition of seedling row centerlines.Firstly,the image database of rice seedling rows under different light conditions(sunny and cloudy days),different weed density distributions and seedling growth status was established during the weed germination period;and then the object detection model based on Faster RCNN network was adopted to detect the positions of rice seedlings;finally,the proposed Hough transform algorithm based on the feature point neighborhood was used to recognize the center line of the seedling row.Experiments indicated that the proposed method had an average accuracy of 92%on the test set,and an average recognition accuracy of seedling rows less than 0.5°under high and low weed density distributions.It was not sensitive to isolated weed noise and light changes,and can also accurately recognize seedling rows with large curvatures.Therefore,the proposed method had good robustness and recognition accuracy.

关 键 词：水稻秧苗行检测 图像数据库 Faster RCNN网络 特征点邻域 HOUGH变换 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]