基于机器视觉的工件自动分拣系统设计

Design of Automatic Sorting System for Workpieces Based on Machine Vision

[目的]机械臂自动分拣系统的研究主要集中在机械臂逆运动学求解和目标检测等方面,缺乏系统的整体设计。[方法]以工业机器人ROS系统为中心,既可以搭建由摄像头、机械臂、树莓派系统、HSV颜色识别、机械臂抓取等软硬件组成的基于ROS话题框架的自动分拣系统;也可以搭建由摄像头、机械臂、树莓派系统、PC上位机、YOLOv5目标检测、机械臂抓取等软硬件组成的基于ROS服务框架的自动分拣系统。同时,分别使用YOLOv5、YOLOv7、YOLOv8等算法在同一台服务器上进行模型训练,训练轮次为100轮,并对比了不同检测算法性能。[结果]采用HSV颜色识别的机器视觉系统响应速度快,但对光照环境、工件颜色要求高,系统适应性差;而YOLOv5目标检测识别率高、精度好,具有普遍性,对工件要求低,适合于一般工作场合,但识别速度有待提高。[结论]采用ROS系统部署分拣系统的不同功能模块,可实现工件的集中式识别和分布式抓取,且各节点分工明确,但该系统仍存在深度学习算法的模型对目标识别速度较慢、光照阴影影响识别精度等问题。未来仍需不断优化深度学习、强化学习的算法,提高目标检测的精度和性能,为未来的智能制造、智能生活提供有力支持。

[Objective]The research on the automatic sorting system of robotic arms mainly focuses on solving the inverse kinematics of robotic arms and object detection,lacking overall system design.[Method]With the industrial robot ROS system as the center,an automatic sorting system based on the ROS topic framework can be built,consisting of software and hardware such as cameras,robotic arms,Raspberry Pi systems,HSV color recognition and robotic arm grasping;an automatic sorting system based on the ROS service framework can also be built,consisting of software and hardware such as cameras,robotic arms,Raspberry Pi systems,PC upper computers,YOLOv5 object detection and robotic arm grasping.At the same time,YOLOv5,YOLOv7,YOLOv8 and other algorithms were used to train the model on the same server for 100 rounds,and the performance of different detection algorithms was compared.[Result]The machine vision system using HSV color recognition has a fast response speed,but has high requirements for lighting environment and workpiece color,and poor system adaptability.YOLOv5 has high target detection recognition rate,good accuracy,universality,low requirements for workpieces,and is suitable for general work environments,but the recognition speed needs to be improved.[Conclusion]By deploying different functional modules of the sorting system using the ROS system,centralized recognition and distributed grabbing of workpieces can be achieved,and each node has clear division of labor.However,the system still faces problems such as slow target recognition speed caused by deep learning algorithms and the impact of lighting shadows on recognition accuracy.In the future,it is still necessary to continuously optimize deep learning and reinforcement learning algorithms to improve the accuracy and performance of object detection,providing strong support for future intelligent manufacturing and intelligent life.