Abstract: Accurate identification and localization of fruit in the natural environment is one of the challenges faced by intelligent picking. Based on binocular stereo vision, this paper proposed a three-dimensional spatial positioning method for outdoor overlapping citrus. Firstly, the contours of overlapping citrus fruits from the left  and right camera  were extracted and Gaussian smoothing was performed. Through the curvature analysis, the abnormal contour pixel points were found. Secondly, the adjacent two abnormal pixel points were connected in turn, and the distance from the pixel points on the line segment to the citrus outline was analyzed, the overlapping citrus contour segmentation was completed at the intersection of two adjacent normal line segments, and the corresponding non-citrus contour pixels were eliminated by searching for the abnormal line segments. Thirdly, the least squares ellipse fitting method was used to reconstruct the contour of the citrus target and obtain the center of the citrus. Finally, according to the binocular line constraint and image similarity, the overlapping citrus center points were matched, and the depth value and three-dimensional space coordinates of the citrus center were calculated based on the parallax principle, and the occlusion relationship of overlapping citrus was obtained by combining the depth values. The results of outdoor experiments showed that the positioning error of this method was 6.38 mm, which met the positioning accuracy requirements of outdoor picking operations of citrus picking robots.

Key words: intelligent picking, binocular stereo vision, overlapping citrus, image similarity, three-dimensional positioning

摘要： 果实的精准识别和定位是智能采摘面临的难题之一。基于双目立体视觉，提出了一种针对户外重叠柑橘的三维空间定位方法。首先，从双目左右图像中提取重叠柑橘果实轮廓并进行高斯平滑，通过曲率分析，找出异常的轮廓像素点；其次，依次连接相邻两个异常像素点，分析该线段上的像素点到柑橘轮廓的距离，在相邻两正常线段的交点处完成重叠柑橘轮廓分割，并通过寻找异常线段剔除对应的非柑橘轮廓像素点；再者，采用最小二乘椭圆拟合方法重建柑橘目标轮廓，并获取柑橘的中心；最后，根据双目极线约束和图像相似度，对重叠柑橘中心点进行匹配，并基于视差原理计算柑橘中心的深度值及三维空间坐标，确定重叠柑橘的遮挡关系。户外实验结果表明,所提出的方法定位误差为6.38 mm，满足柑橘采摘机器人户外采摘作业的定位精度要求。

关键词: 智能采摘, 双目立体视觉, 重叠分割, 图像相似度, 三维空间定位