Abstract: In order to solve the problems of high cost, high labor intensity and slow speed of manual inspection in traditional agricultural production mode, a cherry tomatoes greenhouse inspection robot system was designed, which integrated automatic control technology, machine vision technology and deep learning technology to achieve accurate identification and counting of cherry tomatoes. The robots automatic navigation function is realized by extracting the working path through machine vision and avoiding the obstacle by infrared sensor. An improved YOLOv4-tiny model was proposed to detect cherry tomatoes. The model added 52×52 feature layer and attention mechanism on the basis of YOLOv4-tiny to improve the recognition accuracy of small target detection. The test results show that the inspection robot can accurately work along the predetermined path when traveling at a speed of 0.2m/s. The average recognition accuracy of the improved YOLOv4-tiny-Tomato model for mature and immature cherry tomatoes in the greenhouse environment is 98.7% and 98.0%, and the F1 value reaches 96.71% and 95.8%, respectively. The realtime frame rate reaches 36 frames/s. This research can provide new ideas and technical support for intelligent inspection and fine management in agricultural field.

Key words: greenhouse, inspection robot, automatic navigation, deep learning, object detection, YOLOv4-tiny

摘要： 为解决传统农业生产方式中人工巡检成本高、劳动强度大、速度慢等问题，设计一种圣女果温室巡检机器人系统，融合自动控制技术、机器视觉技术和深度学习技术，实现对圣女果的准确识别和计数。通过机器视觉提取工作路径和红外传感器避障实现机器人自动导航功能；并提出一种改进型YOLOv4-tiny模型对圣女果进行检测，该模型在YOLOv4-tiny的基础上增加52×52的特征层和注意力机制，提高小目标检测的识别精度。试验结果表明，该巡检机器人以0.2m/s的速度行进时可以准确沿着预定路径工作，改进的YOLOv4-tiny-Tomato模型在温室环境中对成熟和未成熟圣女果的平均识别精度分别为98.7%和98.0%，F1值分别达到96.71%和95.8%，实时帧率达到36帧/s。本研究可为农业领域的智能化巡检和精细管理提供新的思路和技术支持。

关键词: 温室大棚, 巡检机器人, 自动导航, 深度学习, 目标检测, YOLOv4-tiny