基于Kinect V3的单株作物自动化三维重建与验证

为高效、精确地对单株作物进行三维重建，以点云方式无损测量作物表型信息，该研究提出一种基于Kinect V3深度相机的三维重建系统。使用步进电机搭建了一个旋转台，并将旋转台面设计成多颜色同心圆，利用同心圆计算平面法向量及圆心两特征信息，用于点云水平校准以及获取点云间的旋转平移矩阵；将Kinect V3采集的多视角点云变换到同一坐标系下，并结合裁剪迭代最近点（Trimmed Iterative Closest Point，TrICP）算法实现了多视角点云的粗配准与精配准，完成了作物三维重建。为检验该研究的三维重建效果，选取菜心、黄瓜苗为试验对象，与多视图立体视觉-运动恢复结构（Multi-View Stereo and Structure From Motion，MVS-SFM）算法重建点云进行对比，并提取叶面积参数与人工测量值进行比较。结果表明，两种方法下重建后的菜心点云间平均距离误差为0.59 cm，黄瓜苗点云间平均距离误差为0.67 cm，具有较高的相似度，而相较于MVS-SFM算法，该研究提出的方法的重建速度提高了约90%；该研究提出的方法所重建点云，菜心叶面积提取与标准参考值相对误差均值为5.88%，均方根误差为3.83 cm2，黄瓜苗叶面积提取与标准参考值相对误差均值别为6.50%，均方根误差为2.08 cm2，都显现出较高的准确性。该研究提出的方法能对单株作物进行快速三维重建，能有效提取叶面积参数，可为作物育种、栽培和农业生产提供高效技术手段和数据支持。

Automatic 3D reconstruction and verification of an individual crop using Kinect V3

Abstract: 3D point clouds can be expected to acquire plant phenotypic traits at present. The damage to crops can be reduced significantly, compared with the traditional manual direct contact measurement. The measurement errors caused by the occlusion between leaves and leaf curl in the two-dimensional images can also be avoided. In this study, a 3D reconstruction system was proposed using Kinect V3, in order to efficiently and accurately reconstruct 3D models of individual crops for the nondestructive measurements of crop phenotype traits by the point cloud. A turntable was also built using a stepping motor. The table surface was designed into multi-color concentric circles. The plane normal vector and central point of the turntable were calculated automatically using concentric circles, which were used for the horizontal alignment of the point cloud and for the calculation of the transformation matrix between multi-view point clouds. The point clouds of crops captured at the multiple view angles by Kinect V3 were transformed to the same coordinate system for the coarse registration using the transformation matrix. Then, the Trimmed Iterative Closest Point (TrICP) was used for the precise registration, thereby completing the 3D reconstruction of individual crops. More importantly, Chinese flowering cabbage and cucumber seedlings were selected as the experimental objects for the 3D reconstruction. Firstly, the reference point clouds were reconstructed by the multi-view stereo and structure from motion (MVS-SFM). The reason was that the MVS-SFM previously presented sufficient accuracy for the crop 3D reconstruction and high-throughput crop phenotyping analysis. A counting was performed on the distribution frequency for the set of distances between the reconstructed and reference point cloud. The results show that the reconstruction speed of the proposed model was improved by about 90%, compared with the MVS-SFM. In the Chinese flowering cabbage, the average error of distance between the reconstructed and the reference point clouds was 0.59 cm; while 64.45% and 90.63% of the distances sets were less than 0.5 and 1.0 cm, respectively. In the cucumber seedlings, the average error of distance was 0.67cm; while 60.09% and 85.45% of the distances sets were less than 0.5 and 1.0 cm, respectively. Both of the groups showed a high level of overlap. Secondly, the single leaf area was extracted by calculating the area of the surface mesh model that was reconstructed using the Delaunay triangular meshing algorithm. The Root Mean Square Error (RMSE) and the average relative error of the Chinese flowering cabbage leaf area were 3.83 cm2 and 5.88%, respectively, compared with the manual measurement using a leaf area meter. The RMSE and the average relative error of cucumber seedling leaf area were 2.08 cm2 and 6.50%, respectively. Both of the groups showed high correlation and accuracy. In addition, the Kinect V3 was compared with the predecessor, Kinect v2, indicating sufficient accuracy for the crop 3D reconstruction. The results show that the Kinect V3 can be used to capture much denser point clouds than the Kinect V2, indicating the high accuracy of crop 3D reconstruction and extraction of leaf area. The proposed model can be expected to quickly reconstruct the individual crops and then effectively extract the leaf area parameters, indicating that it can provide efficient technical tools and data support for crop breeding, cultivation, and agricultural production.