Abstract: In order to solve the current agricultural picking robot problem of target recognizing and localizing difficulties, this paper uses an improved YOLOv3 algorithm combined with 3D vision technology to achieve accurate recognition and localization of the target based on the original agricultural picking robot, and completes the transformation between the target coordinate system and the robot coordinate system by calibration. The performance of improved YOLOv3 algorithm is analyzed through experiments, and compared with the previous YOLOv3 algorithm, Fast RCNN algorithm and Faster RCNN algorithm. The research shows that improved YOLOv3 algorithm and 3D vision have higher recognizing accuracy and localizing accuracy, and the recognizing accuracy is increased by 5.5%, 9% and 1.4% respectively, and the maximum localizing errors were reduced by 0.69, 0.44 and 0.28 mm respectively. It can better complete the followup picking work, which has important reference value for the development of agricultural robot.

Key words: YOLOv3, vision, picking robot, recognition and localization

摘要： 为解决目前农业采摘机器人目标难以识别与定位的问题，在原有农业采摘机器人的基础上，提出一种改进YOLOv3算法和3D视觉技术相结合的方法，实现目标的准确识别和精准定位，并利用标定完成目标坐标系和机器人坐标系的转换。通过试验分析改进YOLOv3算法的性能，并与之前的YOLOv3算法、Fast RCNN算法和Faster RCNN算法进行综合比较，研究表明所采用的改进YOLOv3算法和3D视觉具有较高的识别准确度和定位精度，识别准确率分别提高55%、9%、1.4%，最大定位误差分别降低0.69、0.44、0.28 mm，可以较好地完成后续采摘工作，对于农业机器人的发展具有重要的参考价值。

关键词: YOLOv3, 视觉, 采摘机器人, 识别与定位