基于运动恢复结构的玉米植株三维重建与性状提取

针对传统的玉米植株性状测量方法存在主观性强、劳动强度大、有损伤等问题,提出了基于运动恢复结构(Structure from motion,SfM)的户外玉米植株三维重建方法,并提取了株高、单株最小包围盒体积、茎粗、叶面积、叶片数、叶夹角等11个性状参数。采用前期研制的小车,在户外采集不同视角下的玉米植株图像(采集间距为5~6 cm),基于SfM算法获取玉米植株三维点云;运用直通滤波、圆柱拟合和条件欧氏聚类算法自动分割单株、茎秆和叶片等点云数据,基于距离最值遍历、三角面片化等算法实现株高、茎粗、叶面积等11个性状的准确、无损测量。结果表明,与人工测量值相比,测得的株高、茎粗和叶面积的平均绝对百分比误差分别为3.163%、4.760%和19.102%,均方根误差分别为3.557 cm、1.540 mm、48.163 cm2,决定系数分别为0.970、0.842、0.901。研究表明,本文方法适用于作物表型户外测量,为表型研究提供了一种新的作物表型户外测量方法,同时还证明,株高和单株最小包围盒体积可以显著区分低地上部生物量玉米植株和高地上部生物量玉米植株。

Three-dimensional Maize Plants Reconstruction and Traits Extraction Based on Structure from Motion

Maize is one of the most widely distributed crops in the world,ranking third only to wheat and rice.The plant height,stalk diameter and leaf area of maize are closely related to its yield,the leaf projection area and leaf stem angle have an direct effect on utilization of light energy to maize plants,the number of leaves is the indicator of the overground part biomass,the parameters such as minimum enveloping box volume of single leaf,leaf perimeter,leaf projection width,leaf projection length and so on directly affect the spatial distribution of leaves,therefore,dynamic monitoring of these traits is particularly important.However,the traditional measurement of these traits is time-consuming,costly,subjective and destructive.To achieve the dynamic,rapid,accurate and non-destructive outdoor measurement of maize plant height,stalk diameter,leaf area,the number of leaves,leaf stem angle and so on,three-dimensional(3D)models of tassel stage maize plants were reconstructed by using structure from motion(SfM)algorithm.An autonomous crawler phenotyping robot was used for acquiring multi-view maize plants images along the maize crop rows outdoors.The robot could work continuously four hours at speed of 0.1 m/s and would acquire about 700 stable images for a single camera.The 3D point cloud data were obtained using the multi-view images in the Visual SFM software.The 3D point cloud data were preprocessed and some morphological traits such as maize plant height,minimum enveloping box volume of single plant,stalk diameter,the number of leaves,leaf perimeter,leaf area,minimum enveloping box volume of single leaf,leaf projection area,leaf projection width,leaf projection length and leaf stem angle were extracted in the Visual Studio 2013 plus PCL platform.Compared with the manual measurement,the mean absolute percentage errors(MAPE)for plant height,stalk diameter and leaf area were 3.163%,4.760%and 19.102%,respectively.The root mean square error(RMSE)for plant height,stalk diameter and leaf area were 3.557 cm,1.540 mm and 48.163 cm2,respectively.The R2 for plant height,stalk diameter and leaf area were 0.970,0.842 and 0.901,respectively.The results showed that 3D reconstruction method based on SfM algorithm was suitable for outdoor measurement.In addition,the maize plants were divided into low overground part biomass maize and high overground part biomass maize by the fresh weight of the overground part plant,meanwhile,the plant trait such as height,minimum enveloping box volume of single plant,stalk diameter and the number of leaves were extracted with segmented point cloud data to calculate the P value by single factor analysis of variance.The measured P values were 0.0003,0.0004,0.3170 and 0.2415,respectively,and the results proved that the traits of plant height and minimum enveloping box volume of single plant were able to distinguish the low overground part biomass maize and high overground part biomass maize evidently.The research result provided scientific researchers and crop breeders a new phenotyping method for measuring crop traits to some extent.