基于机器视觉的幼苗自动嫁接参数提取

为提高果蔬嫁接机器人的自动化水平，该文提出一种基于机器视觉用椭圆拟合的方法恢复幼苗叶面并提取用于机器人自动嫁接的参数的方法。俯视采集幼苗图像，提取叶面轮廓并根据轮廓上的拐点对组合相同叶面上的轮廓弧段。应用椭圆拟合的方法参数化叶面形状，提取幼苗的叶面参数，包括生长方向、生长点和叶面面积。再由生长点准确定位培育幼苗的穴孔位置，从而为砧穗配对和取苗定位提供依据。试验结果表明提出的算法能够克服叶面相互遮挡的问题，幼苗识别且定位的成功率达到97.5%，能满足嫁接机器人自动作业的要求。

Vision-based parameters extraction of seedlings for grafting robot

Abstract: To achieve full automation of a grafting robot for fruits and vegetables, this paper presented a machine vision system for restoring the cotyledons of seedlings and extracting their parameters by ellipse fitting. Overlooking images of seedlings were captured by a gray camera. After doing a fast median filter, bright areas composed of cotyledons were segmented by an auto-threshold binarization algorithm with Otsu. Moreover, their contours could be easily found by an edge trace algorithm. Each corner representing the intersection of two different cotyledons was detected by finding the local maximum of the curvature in the contours. Under the constraints of distance and arc length, two corners which belong to the same two cotyledons were made a pair. Thus, all contour segments belonging to the same cotyledon could be retrieved. To restore the shape of the cotyledon, its contour was parameterized by a fitting ellipse. Through combining two parameterized cotyledons, the parameters of seedlings including growth direction, growth point, and size were extracted. Furthermore, the cells in which the seedlings were planted were determined by arraying all growth points. In a grafting operation, the growth direction and growth point can be used to assist the manipulator to fetch the seedling accurately, and the size of a cotyledon can provide some information for making the best correspondences between rootstocks and scions. One example showed that the proposed method can achieve good performance even if the cotyledons overlap each other. The errors from an ellipse fitting defined as the least absolute distance from initial points to the ellipse were calculated. Moreover, their means and standard derivation were mostly near 0.5 pixels, which indicated that the ellipse can represent the shape of the cotyledon well. Exceptionally, several ellipses with large errors were not accurate due to the fuzzy contour. The reason is that the segmented bright area was mixed in with some non-cotyledon region. The final test showed that 461 seedlings were identified and positioned in all 473 seedlings, and its rate reached 97.5%, which meets the requirement of robot grafting. There are two main reasons why some seedlings were missed. First, the bright stem and the cotyledon may be overlapped in the overlooking image because of the bending stem, and may bring in an additional corner of the contour. Secondly, when the place the seedling is living is kept away from the center of tray or the two neighbor cotyledons are too close, the cell where the seedling was planted cannot be found.