遮挡条件下多视角甜椒果实点云三维重构方法

为进行表型原位自动化测量，实现甜椒数字化育种和管理，针对原位果实表型测量中的目标遮挡问题，提出一种多视角甜椒果实点云的三维重构方法。通过虚拟叶片的方法，创建增强数据集，建立基于YOLO v5算法的甜椒果实识别模型，实现对不同遮挡程度果实的识别，同时，构建考虑果实位置与遮挡程度的果实表型采集算法，实现多视角的果实三维数据采集。最后，配准甜椒果实三维点云，提取甜椒表型参数，并通过温室甜椒果实表型，对点云重构方法的有效性进行验证。相较手动测量数据，果实果宽平均相对误差为1.72%，果高平均相对误差为1.60%。试验结果表明，本文所提出的甜椒原位表型点云重构方法，可为遮挡条件下作物表型提供有效的解决思路和可行方法。

Multi Perspective Point Cloud Reconstruction Method for Sweet Pepper Fruit under Occlusion Conditions

The in-situ phenotype of sweet pepper is an important reference indicator for fruit breeding and management. Automated measurement of phenotype in-situ through phenotype collection robots is one of the effective ways for digital breeding and management of sweet pepper. However, fruit occlusion during the measurement process seriously affects the success rate of detection. Therefore, a three-dimensional reconstruction method for multi view sweet pepper fruit point cloud was proposed to address the problem of target occlusion in in-situ fruit phenotype measurement. By using the method of virtual leaves, an enhanced dataset was created, and a sweet pepper fruit recognition model based on YOLO v5 algorithm was established to recognize fruits with different degrees of occlusion. At the same time, a fruit phenotype collection algorithm considering fruit position and occlusion degree was constructed to achieve multi view three dimensional data collection of fruits. Finally, the three-dimensional point cloud of sweet pepper fruit was registered, the phenotype parameters of sweet pepper was extracted, and the effectiveness of the point cloud reconstruction method was validated through the greenhouse sweet pepper fruit phenotype. Compared with manual measurement data, the average relative error of fruit width was 1.72%, and the average relative error of fruit height was 1.60%. The experimental results indicated that the in-situ phenotype point cloud reconstruction method proposed for sweet pepper can provide effective solutions and feasible methods for crop phenotypes under occlusion conditions.