基于树莓派平台的柑橘智能识别与采摘

Design of citrus identification and picking based on the Raspberry PI platform

【目的】基于树莓派平台,设计轻便式的柑橘智能识别与机器人采摘,为柑橘采摘减轻劳动力与机械智能化提供理论与实验依据。【方法】制作柑橘数据集,使用深度卷积神经网络(YOLOv4-tiny)验证柑橘数据集的有效性,同时在电脑上离线训练好YOLOv4-tiny柑橘识别模型;基于树莓派平台,嵌入YOLOv4-tiny模型,设计柑橘智能识别系统;建立五自由度机械臂运动学模型与轨迹规划函数,嵌入树莓派平台。使用树莓派开发板、一个摄像头以及小型五自由度机器人,手动启动树莓派中的YOLOv4-tiny模型,系统自动完成柑橘识别与采摘,具体步骤为:调用OpenCV开启摄像头,将摄像头采集到的视频转换为Im格式,并输入YOLOv4-tiny模型中,模型识别出待采摘柑橘的位置,机械臂再进行实际采摘。【结果】深度卷积神经网络(YOLOv4-tiny)模型对橘子二维坐标识别率为98%,嵌入树莓派后,结合目标到相机距离相似三角形计算法,能准确识别柑橘位置三维坐标(-3.2,19.7,16);驱动机械臂到达位置坐标后,再令夹持器夹紧并摘取柑橘,先后共调试80次,得到机械臂运动最佳时间为1000ms、夹持器末端采摘动作的最佳运动时间为500ms,末端夹持器摘取的平均角度为-60度,这样采摘时速度合适也不会发生抖动;进行了20次采摘测试,平均成功摘取率达到了90%以上。【结论】基于树莓派平台设计的柑橘深度识别与采摘系统,能实现柑橘三维位置坐标的识别并准确采摘,为机器人柑橘采摘的智能化提供了一定的理论基础与实验借鉴。

【Objective】Based on the Raspberry PI platform,a portable citrus intelligent recognition and robot picking system were designed to provide a theoretical and experimental basis for reducing labor force and mechanical intelligence in citrus picking.【Method】The citrus data set was made,and the deep convolutional neural network(YOLOv4-tiny)was used to verify the validity of the data set.Based on the Raspberry PI platform combined with YOLOv4-tiny,a citrus intelligent recognition system was designed.The kinematic model and trajectory planning function of the five-degree-of-freedom manipulator were established,and the Raspberry PI development board,a camera and a small five-degree-of-freedom robot were used for the intelligent identification and picking of citruses.【Result】The recognition rate of the two-dimensional coordinates of citruses by the deep convolutional neural network(YOLOv4-tiny)model was 98%.After embedding Raspberry PI,combined with the similar triangle calculation method of the distance between the target and camera,the three-dimensional coordinates of citruses can be accurately identified(-3.2,19.7,16).After driving the mechanical arm to the position coordinate,the gripper clamp was made and the citrus can be picked up.After a total of 80 times of adjustment,the best time of mechanical arm movement was 1000 ms,the best time of picking the end of the gripper was 500 ms,and the average angle of picking the end of the gripper was-60 degrees,so that the picking speed was appropriate without jitter.20 harvest tests were conducted,and the average successful capture rate reached more than 90%.【Conclusion】The depth recognition and picking system based on the Raspberry PI platform can realize the recognition and accurate picking of three-dimensional coordinates of citruses,which provides a certain theoretical basis and experimental reference for intelligent citrus picking.