温室高架栽培草莓空间姿态识别与采摘点定位方法

采摘目标空间位姿信息缺失和目标定位精度低是影响草莓采摘机器人采摘效果的关键因素之一。为此，本文首先设计了基于颜色信息和卷积神经网络的草莓图像目标定位与分割以及目标点云分割模型；其次，实现了基于图像的草莓可采摘性和遮挡程度识别模型；最后，设计了草莓空间定位和姿态估计模型并实现草莓采摘点定位方法。基于本文方法对完整草莓位姿估计平均误差为4.03%，对遮挡草莓位姿估计平均误差为9.06%，采摘定位综合误差为2.3mm。在实际采摘实验中，采摘成功率为92.6%，平均每个草莓的计算耗时约为92ms，单个草莓采摘动作的执行平均耗时约为5.7s。实验结果表明：本文提出的方法可在温室条件下较准确地估计草莓空间位姿和采摘点，为草莓采摘机器人提供有效的目标定位信息，有效满足实际采摘场景下的需求。

Spatial Posture Recognition and Picking Point Location Method for Greenhouse Raised-frame Strawberry Cultivation

The lack of spatial positional information of picking targets and low target localization accuracy are one of the key problems that limit the picking effect of strawberry picking robots. To address these problems, a target localization and segmentation model was firstly designed based on color information and convolutional neural network for strawberry image and target point cloud segmentation;secondly, an image-based strawberry pickability and obscuration recognition model was implemented;finally, a strawberry spatial localization and pose estimation model was designed and a strawberry picking point localization method was implemented. Based on this method, the estimation error of intact strawberry position was 4.03%, the estimation error of obscured strawberry position was 9.06%, and the comprehensive error of picking position was 2.3mm. In the actual picking experiment, the picking success rate was 92.6%, the average calculation time of each strawberry was about 92ms, and the average execution time of single strawberry picking action was about 5.7s. The experimental results can provide effective target localization information for strawberry picking robots, which can effectively meet the needs of actual picking scenarios.