基于三维点云的黄瓜叶片分割与表型参数提取方法

Cucumber Leaf Segmentation and Phenotype Extraction Method Based on Three-dimensional Point Cloud

摘要:

自动获取植株冠层表型形状对黄瓜育种和科学栽培至关重要。由于当前三维点云处理技术难以在黄瓜植株点云上对茎叶进行有效分离，分割准确率和效率较低。本文提出了一种改进的区域生长分割算法，并对分割后叶片进行表型提取。首先通过深度相机从4个角度采集黄瓜点云数据，在统计滤波和颜色滤波去除背景噪声以及离群点的基础上，基于旋转轴和广义最近点迭代（Generalized nearest point iterative，GICP）算法对点云进行配准获取完整黄瓜植株点云；使用体素和移动最小二乘算法（Moving lest squares，MLS）对区域生长算法进行改进，实现茎叶分离与叶片分割；分割后叶片点云自动提取叶片数量、叶面积、叶长、叶宽、叶周长表型参数。实验结果表明，与传统区域生长算法相比，改进区域生长算法可以精准地分割出单个叶片，对移栽15 d的准确率平均提升12.5个百分点，对移栽60 d的准确率平均提升22.5个百分点。叶面积、叶长、叶宽、叶周长4个参数与真实测量值相比决定系数R2分别为0.96、0.93、0.93、0.94，均方根误差（RMSE）分别为12.69 cm2、0.93 cm、0.98 cm、2.27 cm。本文提出的方法能够从单株黄瓜点云中高效地分割出单个叶片点云，并准确地计算相关表型性状，为温室黄瓜高通量自动化表型测量提供有力的技术支持。

Abstract:

Automatic acquisition of plant canopy phenotypic shape is essential for seed selection and scientific cultivation of cucumber varieties. Segmentation accuracy and efficiency are low due to the difficulty of current 3D point cloud processing techniques to perform effective separation of stems and leaves on cucumber plant point clouds. Aiming to address this problem, an improved algorithm for regional growth segmentation and phenotype extraction was proposed by segmented leaves. Firstly, the point cloud data of cucumber was collected from four angles by depth camera, and on the basis of statistical filtering and color filtering to remove the background noise as well as outliers, the complete cucumber plant point cloud by aligning the point cloud-based on rotary axis and generalized nearest point iterative algorithm (GICP), and then the region growth algorithm was improved by using voxel-based and moving least squares algorithms (MLS) to realize the separation of stems and leaves and the segmentation of leaves;finally, phenotypic parameters such as number of leaves, leaf area, leaf length, leaf width, leaf circumference were automatically extracted from the segmented leaf point cloud. The experimental results showed that individual leaves could be accurately segmented by the improved zone-growth algorithm compared with the traditional zone-growth algorithm, with an average increase in accuracy of 12.5 percentage points for 15 d of transplanting and 22.5 percentage points for 60 d of transplanting. The coefficient of determination R2 for the four parameters of leaf area, leaf length, leaf width, and leaf circumference were 0.96, 0.93, 0.93, and 0.94, respectively, and the root-mean-square error RMSE was 12.69 cm2, 0.93 cm, 0.98 cm, and 2.27 cm, respectively, compared with the true measurements. Therefore, the proposed method can efficiently segment individual leaf point clouds from a single cucumber point cloud and accurately calculate related phenotypic traits, providing strong technical support for high-throughput automated phenotypic measurements in greenhouse cucumbers.