基于神经辐射场的RGB图像点云重建多肉植物及尺寸测量研究

Point Cloud Reconstruction of Succulent Plants Based on Neural Radiance Fields RGB Image and Dimensional Measurement

以多肉植物盆栽为研究对象,使用手持式RGB相机采集11个多肉植物盆栽的视频数据,通过将视频转换为图像帧、选取优质清晰图像帧、计算相机位姿得到含丰富信息的RGB图像数据。提出一种改进神经辐射场的多肉植物三维重建方法,根据实际场景提出新的射线采样策略,同时引入改进的图像修复模块与隐式模型重建点云方法,并根据点云重建结果提取多肉植株的叶片数、株高、冠围、凸包体积、叶长、叶宽和叶色共7个表型参数。最后选取具有代表性、易测量的叶片数、株高、冠围、叶长和叶宽5个表型参数进行精度评估与误差原因分析,平均绝对百分比误差(MAPE)分别为2.32%、3.95%、4.95%、5.59%和9.55%,均方根误差(RMSE)分别为0.86片和1.95、17.54、1.87、1.27 mm,决定系数(R^(2))分别为0.99、0.99、0.86、0.91和0.89。精度评估结果表明,所提取的表型参数能够准确、高效地反映多肉植株生长状态,充分发挥RGB图像新视角合成技术、图像处理技术与三维点云重建技术的优势,实现多肉植株盆栽的表型参数高精度、非破坏性提取,能够为多肉植物的种植和养育以及为非固定、多视角的RGB数据获取研究提供重要的技术支持。

Focusing on potted succulent plants,handheld RGB cameras were utilized to collect video data of 11 potted succulent plants.By converting videos into image frames,high-quality clear frames were selected,and camera poses were calculated,and containing rich information RGB image data was obtained.An improved method for three-dimensional reconstruction of succulent plants based on NeRF was proposed.A new ray sampling strategy tailored to actual scenes was introduced,along with an enhanced image restoration module and an implicit model for point cloud reconstruction.Seven phenotypic parameters of succulent plants were extracted from the point cloud reconstruction results,including leaf count,plant height,crown circumference,convex hull volume,leaf length,leaf width,and leaf color.Finally,a precision assessment and error analysis were conducted on five representative and easily measurable phenotypic parameters:leaf count,plant height,crown circumference,leaf length,and leaf width.The mean absolute percentage error(MAPE)for these parameters was respectively 2.32%,3.95%,4.95%,5.59%,and 9.55%,and the root mean square error(RMSE)was respectively 0.86 leaves and 1.95 mm,17.54 mm,1.87 mm,1.27 mm,with respective R^(2) values of 0.99,0.99,0.86,0.91,and 0.89.The results of precision assessment indicated that the extracted phenotypic parameters can accurately and efficiently reflect the growth status of succulent plants.By leveraging advantages in RGB image synthesis technology,image processing,and 3D point cloud reconstruction,non-destructive extraction of phenotypic parameters for potted succulent plants was achieved with high precision.The research result can provide important technical support for succulent plant cultivation and nurturing,as well as for studies involving non-fixed,multi-perspective RGB data acquisition.