基于逆运动学降维求解与YOLO v4的果实采摘系统研究

为提高采摘设备的执行效率，采用六自由度机械臂、树莓派、Android手机端和服务器设计了一种智能果实采摘系统，该系统可自动识别不同种类的水果，并实现自动采摘，可通过手机端远程控制采摘设备的起始和停止，并远程查看实时采摘视频。提出通过降低自由度和使用二维坐标系来实现三维坐标系中机械臂逆运动学的求解过程，从而避免了大量的矩阵运算，使机械臂逆运动学求解过程更加简捷。利用Matlab中的Robotic Toolbox进行机械臂三维建模仿真，验证了降维求解的可行性。在果实采摘流程中，为了使机械臂运动轨迹更加稳定与协调，采用五项式插值法对机械臂进行运动轨迹规划控制。基于Darknet深度学习框架的YOLO v4目标检测识别算法进行果实目标检测和像素定位，在Ubuntu 19.10操作系统中使用2000幅图像作为训练集，分别对不同种类的果实进行识别模型训练，在GPU环境下进行测试，结果表明，每种果实识别的准确率均在94%以上，单次果实采摘的时间约为17s。经过实际测试，该系统具有良好的稳定性、实时性以及对果实采摘的准确性。

Design of Fruit Picking System Based on Inverse Kinematics Dimension Reduction and YOLO v4

In order to improve the execution efficiency of picking equipment, a six DOF robot arm, raspberry pie, single chip microcomputer, Android mobile terminal and server were used to build an intelligent fruit picking experimental system. It can automatically identify different kinds of fruits and realize automatic picking. Users can control the picking equipment and view the real-time picking video through the mobile terminal remotely. The solution process of inverse kinematics of manipulator in the three-dimensional coordinate system was realized by reducing the degree of freedom and using two-dimensional coordinate system, so as to avoid a lot of matrix operation and make the process of inverse kinematics solution of manipulator more simple. The Robotic Toolbox in Matlab was used to carry out 3D modeling and simulation of the manipulator, and the feasibility of dimension reduction was verified. In the fruit picking process, in order to make the trajectory of the manipulator more stable and coordinated, the pentanomic interpolation method was used to plan and control the trajectory of the manipulator. The object detection and pixel location of fruit was based on the YOLO v4 artificial intelligence recognition algorithm of Darknet deep learning framework. Using GPU training in the Ubuntu 19.10 operating system, totally 2000 images were used as a training set to train the recognition model of different kinds of fruits. The accuracy rate of each fruit recognition was more than 94%, the time for a single fruit picking about 17s. After the actual test, the system had good stability, real-time performance and the accuracy of fruit picking.