基于神经辐射场的苗期作物三维建模和表型参数获取

苗期作物三维结构的精准高效重建是获取表型信息的重要基础。传统的三维重建大多基于运动恢复结构-多视图立体视觉（Structure from motion and multi-view stereo，SFM-MVS）算法，计算成本高，难以满足快速获取表型参数的需求。本研究提出一种基于神经辐射场（Neural radiance fields，NeRF）的苗期作物三维建模和表型参数获取系统，利用手机获取不同视角下的RGB影像，通过NeRF算法完成三维模型的构建。在此基础上，利用点云库（Point cloud library，PCL）中的直线拟合和区域生长等算法自动分割植株，并采用距离最值遍历、圆拟合和三角面片化等算法实现了精准测量植株的株高、茎粗和叶面积等表型参数。为评估该方法的重建效率和表型参数测量精度，本研究分别选取辣椒、番茄、草莓和绿萝的苗期植株作为试验对象，对比NeRF算法与SFM-MVS算法的重建结果。结果表明，以SFM-MVS方法重建点云为基准，NeRF方法重建的各植株点云点对距离均方根误差仅为0.128~0.395cm，两者重建质量较接近，但在重建速度方面，本文研究方法相比于SFM-MVS方法平均重建速度提高700%。此外，该方法提取辣椒苗株高、茎粗决定系数（R2）分别为0.971和0.907，均方根误差（RMSE）分别为0.86cm和0.017cm，对各苗期植株叶面积提取的R2为0.909~0.935，RMSE为0.75 ~3.22cm2，具有较高的测量精度。本研究提出的方法可以显著提高三维重建和表型参数获取效率，从而为作物育种选苗提供更为高效的技术手段。

Three-dimensional Reconstruction and Phenotype Parameters Acquisition of Seeding Vegetables Based on Neural Radiance Fields

Accurate and efficient reconstruction of seedling crop structures is crucial for obtaining phenotype parameters. The traditional method for 3D reconstruction based on the structure from motion and multi-view stereo (SFM-MVS) algorithm, which had high reconstruction accuracy and high computional cost. It was difficult to meet the demand for rapid acquisition of phenotype parameters. A system for acquiring phenotype parameters and creating 3D models of seedling crops was proposed by using neural radiance fields (NeRF). The system utilized smart phone to capture RGB images of the objects from various viewpoints and constructed the 3D model through the NeRF algorithm. The algorithms of line fitting and region growing in point cloud library (PCL) were used to automatically segment the plants. Additionally, the algorithms of distance-minimum traversal, circle fitting, and triangulation were used to measure phenotype parameters such as plant height, stem diameter, and leaf area. To assess the reconstruction efficiency and accuracy of phenotype parameter measurement, seedling plants of pepper, tomato, strawberry and epipremnum aureum were selected as subjects. The reconstruction results were compared by using the NeRF and the SFM-MVS algorithm. The results indicated that both methods were capable of achieving superior reconstruction outcomes. The root mean square errors of the point-to-point distances of each seedlings were only 0.128cm to 0.359cm. But in terms of speed, this method improved the reconstruction speed by an average of 700% compared with the SFM-MVS method. The method used to extract plant height and stem diameter of chili pepper seedlings had a coefficient of determination (R2) of 0.971 and 0.907, respectively. The root mean square error (RMSE) was 0.86cm and 0.017cm, respectively. The R2 of the leaf area extracted from the plants at seedling stage ranged from 0.909 to 0.935, and the RMSE ranged from 0.75cm2 to 3.22cm2, indicating a high level of accuracy in measurement. The proposed method can significantly speed up 3D reconstruction and acquisition of phenotype parameters. This would provide a more efficient technical means for vegetable breeding and seedling selection.