Authors

Abstract

To address the issue of 3D reconstruction information loss caused by occlusion during single-view camera acquisition of crop phenotypic parameters, this study proposes a detection method for tomato plant phenotypic parameters based on multi-view 3D point cloud reconstruction. The Kinect 2.0 sensor was employed to acquire point cloud data of tomato plants from three different viewpoints. Background noise was effectively removed using a combination of Conditional Filtering and Statistical Outlier Removal methods. By extracting surface normal features and calculating Fast Point Feature Histograms (FPFH), the Sample Consensus Initial Alignment (SAC-IA) and Iterative Closest Point (ICP) algorithms were utilized to accomplish coarse and accurate registration of the point clouds, respectively, ultimately achieving 3D reconstruction. Experimental results demonstrated that the reconstructed 3D model of the tomato plant was clear in outline and complete in structure. For the phenotypic parameters of plant height, canopy width, and leaf angle, the coefficients of determination (R²) between the calculated and manually measured values were 0.98, 0.94, and 0.89, respectively, with Root Mean Square Errors (RMSE) of 0.75 cm, 1.10 cm, and 4.43 °. Compared to single-view measurements, the accuracy of plant height and maximum canopy width derived from multi-view reconstruction increased by 15.31% and 13.12%, respectively. This method provides technical support for the rapid and accurate extraction of phenotypic parameters in tomato plants.

Abstract in Chinese

针对摄像头单视角采集作物表型参数过程中因遮挡导致三维重建信息缺失的问题，提出一种基于多视角三维点云重建的番茄植株表型参数检测方法。采用Kinect 2.0传感器获取三个不同视角番茄植株的点云数据，利用条件滤波与统计离群点移除方法去除背景噪声。通过提取表面法线特征并计算快速点特征直方图（FPFH），利用采样一致性初始配准（SAC-IA）与迭代最近点（ICP）算法完成点云的粗配准与精配准，实现三维重建。实验结果表明，重建后的番茄植株三维模型外形轮廓清晰、结构完整。植株株高、最大冠幅和叶夹角等表型参数，计算值与实测值的决定系数R²分别为0.98、0.94和0.89，均方根误差RMSE分别为0.75 cm、1.10 cm和4.43°。与单视角对比，多视角重建的株高和最大冠幅的精度分别提高了15.31%和13.12%。该方法为番茄植株表型参数的快速、精准提取，提供了技术支持。