【作者】 程志强；

【导师】 方圣辉；

【作者基本信息】 武汉大学 ， 摄影测量与遥感， 2023， 硕士

【摘要】 在现代农业中,水稻对于满足全球人口的粮食需求而言至关重要。因此,水稻表型研究成为了农业科学中重要的领域之一,它涉及到了水稻的生长发育、生产性状和环境适应性等方面。同时,水稻表型研究也为改良和优化水稻品种、提高水稻产量和质量、实现农业的可持续发展提供了重要参考。当前一些研究关注于传统基于二维图像的水稻表型检测方法,然而二维图像缺乏空间维度信息,难以从中提取全面的水稻表型参数,因此建立水稻三维模型进而提取全面的水稻表型参数具有重要意义。当前植株的三维重建方法主要分为基于传感器的主动视觉法和基于多视角图像的摄影测量方法。相较于基于传感器的主动视觉法,基于多视角图像的三维重建方法具有更高的效率和更低的成本。此外,这类方法不受场景限制且能够获取更多的细节信息。基于多视角图像的三维重建方法主要分为基于特征匹配的三维重建和基于多站摄影的空间雕刻方法,其中空间雕刻以多视角的二值轮廓为输入,通过利用视角交集产生植株的可视化外壳,该方法不依赖于特征匹配的结果且对各类植株均保持有良好的适用性。水稻植株纹理稀疏且叶片间的相互遮挡问题严重,因此难以通过特征匹配完成水稻植株的三维重建。为此,本研究基于空间雕刻方法重建了高精度的水稻三维模型。本研究以盆栽水稻为实验对象,首先利用相机标定原理获取相机和水稻位置,然后使用色彩阈值提取图像序列水稻轮廓,其次利用空间雕刻方法构建水稻三维体素模型,最后使用点云和Alpha-shape方法重构水稻模型。基于水稻三维重建结果,本研究提取并评估了水稻的株高、茎粗、植被覆盖度及体积参数,并基于这些参数完成了对水稻株型的分类以及水稻生物量的估测。实验结果显示,本研究方法的水稻株高、茎粗、单株植被覆盖度的RMSE和MAPE分别为60.34mm、3.95mm、5.32%和9.13%、13.70%、11.89%,均优于现有的基于特征匹配方法的结果。另外,水稻生物量RMSE和MAPE分别为57.9g、17.61%,也均保持在较小范围内。本研究方法可以重建高精度的水稻三维模型,并准确地测算水稻表型参数,为水稻表型研究提供技术支撑,具有较强的可用性。更多还原

【Abstract】 In modern agriculture,rice is crucial for meeting the food needs of the global population.Therefore,rice phenotyping research that provides important references for optimizing rice varieties,increasing rice yield and sustaining the development of agriculture has become one of the important fields in agricultural science,involving rice growth and development,production traits and environmental adaptability.Currently,some studies focus on traditional 2D-image-based rice phenotyping detection methods.Nevertheless,2D-image lacks of sufficient spatial dimension information,making it difficult to extract comprehensive rice phenotyping parameters from 2D-image.Therefore,it is of great significance to establish rice 3D models and extract comprehensive rice phenotyping parameters.As for 3D reconstruction of plants,commonly used methods are mainly divided into sensor-based active vision methods and multi-view images-based photogrammetric methods.Compared to sensor-based active vision methods,multi-view images-based methods take advantages of higher efficiency and lower cost.Additionally,they are unconstrained by scene and able to obtain detailed information.Multi-view images-based 3D reconstruction methods mainly contain two branches,which are feature matching-based 3D reconstruction and space carving method based on multistation photograph.Especially,space carving methods use multi-view binary contours as input and produce plant’s visualized shell by utilizing the intersection of views,which independent of feature matching results and are generally applicable to various plants.Due to rice plants suffer from sparse textures and mutual occlusion between their blades,it is difficult to reconstruct rice 3D models by feature matching methods.For this,our study reconstruct high-precision rice 3D models based on space carving methods.In this study,potted rice plants are used as experimental objects.Firstly,camera calibration principle is utilized to obtain the camera and rice positions,then we divide rice contours from a sequence of images by setting color thresholds.Subsequently,the rice 3D voxel models are constructed by the space carving method.The rice models are reconstructed by means of point cloud and Alpha-shape finally.Furthermore,we also extract and evaluate the height,stem thickness,vegetation coverage,and volume parameters of rice in this study based on the 3D reconstruction results and conduct rice plant type classification and rice biomass estimation based on these extracted parameters.The experimental results show that the RMSE and MAPE of rice height,stem thickness,and single plant vegetation coverage of this study’s method are 60.34mm,3.95mm,5.32%,and 9.13%,13.70%,11.89%,respectively,which are the superior than the results of existing feature matching-based methods.Moreover,the RMSE and MAPE of rice biomass are 57.9g and 17.61%,respectively,which are both contained in a low level.In conclusion,the proposed method can reconstruct high-precision rice 3D models and measure rice phenotyping parameters accurately,providing technical support for rice phenotyping research and having strong availability.更多还原