基于RANSAC拟合点云去噪的苹果采摘位姿构建方法

Apple Picking Pose Establishment Based on Filtering Point-cloud Noise by RANSAC Fitting

摘要:

针对果园环境下果实重叠和光照等因素带入的难滤除点云噪声，导致借助点云构建的采摘位姿精度低的问题，本文提出了一种基于随机采样一致性（Random sample consensus，RANSAC）拟合点云去噪的采摘位姿构建方法。该方法通过RANSAC拟合算法从预处理后的果实点云中检测出多个潜在的球体，并以与点云采集设备垂直距离最短的球体球心作为目标果实的基准设置距离阈值，以便进一步滤除目标果实点云中难滤除的点云噪声，提高目标果实的位姿构建精度。在此基础上，利用最小二乘法对去噪后的点云进行球拟合得到球心坐标，并作为目标果实的采摘位置，然后结合实例分割算法获取的二值化掩膜图像质心三维坐标，构造接近向量作为采摘姿态，完成采摘位姿的构建。重叠果实点云去噪试验表明，本文方法能够有效滤除目标果实中难滤除的点云噪声;位姿构建评估试验结果显示，在室外仿果园环境下采用提出的位姿构建方法，果实定位精度达到15.0mm，相较于直接使用RANSAC拟合球的定位方法，定位精度最大提高28.1%，位置构建稳定性提高76.0%;果园采摘对比试验表明，采用提出的位姿构建方法定位成功率达到70.2%，相较于现有同类方法，定位成功率提高23.4%，采摘成功率提高38.4%。本文提出的方法可为复杂果园环境下的果实位姿准确构建提供参考。

Abstract:

To address the issue of low accuracy in picking pose establishment caused by overlapping fruit and challenging lighting conditions that introduced difficult-to-filter point cloud noise in orchard environments, an accurate method for establishing picking poses based on point cloud denoising using the random sample consensus (RANSAC) algorithm was proposed. Multiple potential spheres were detected from the pre-processed fruit point clouds by using the RANSAC algorithm. The sphere center with the shortest vertical distance to the point cloud capturing device was used to set a distance threshold, which facilitated further noise filtering from the target fruit point clouds and enhanced pose establishment accuracy. Subsequently, the denoised point clouds were spherefitted by using the least squares method to obtain the sphere center coordinates, which defined the precise picking position. Furthermore, by integrating the centroid coordinates from the corresponding binary mask image generated via an instance segmentation algorithm, an approach vector was constructed to determine the harvesting orientation, completing the pose establishment process. Experimental results on overlapping fruit point cloud denoising demonstrated that the proposed method effectively removed challenging point cloud noise from the target fruits. Pose establishment evaluations in an outdoor simulated orchard showed that the proposed method achieved a positioning accuracy of 15.0mm, enhancing the direct RANSAC fitting approach by up to 28.1% in accuracy and 76.0% in stability. Comparative harvesting trials in the orchard confirmed a successful positioning rate of 70.2% by using the proposed approach, which represented an increase of 23.4% over existing methods and a 38.4% improvement in harvesting success. The proposed method offered a robust solution for accurate fruit pose establishment in complex orchard environments.