%0 Journal Article
%T 基于GR-CNN改进的抓取检测算法及仿真<br>Improved Grasping Detection Algorithm and Simulation Based on GR-CNN
%A 娄晓恒
%A 毛鹏军
%A 牛鑫宇
%A 段云涛
%J Modeling and Simulation
%P 5196-5206
%@ 2324-870X
%D 2023
%I Hans Publishing
%R 10.12677/MOS.2023.126473
%X 随着计算机视觉的发展，抓取检测在机器人视觉领域取得了广泛应用，以往的抓取检测方法存在对抓取相关特征学习不够重视、检测精度低、泛化能力差等问题。GR-CNN算法解决了之前这些检测算法存在的一些问题，但是该算法预测的抓取框所处位置和角度有时会存在不准确的情况。针对这一问题，对该算法进行改进优化。通过加入CBAM注意力机制模块使网络主要集中于学习可抓取的信息，从而提高抓取预测的准确性。并在ROS环境下对改进后的算法进行验证，通过Gazebo与MoveIt!联合仿真，使用UR5机械臂对桌面上随机放置的四类物体进行抓取放置任务，平均抓取成功率达到86%，实验结果表明改进后的算法抓取框适应性更强，对规则物体与形状结构复杂的物体都可以有效抓取。<br />
With the development of computer vision, grasp detection has been widely used in the field of robot vision. The previous grasp detection methods have some problems, such as not paying enough at-tention to grasp-related feature learning, low detection accuracy and poor generalization ability. The GR-CNN algorithm solves some problems existing in the previous detection algorithms, but the position and Angle of the grab frame predicted by the algorithm are sometimes inaccurate. To solve this problem, the algorithm is improved and optimized. By adding CBAM attention mechanism module, the network mainly focuses on learning the information that can be captured, so as to im-prove the accuracy of grasping prediction. The improved algorithm was verified in ROS environ-ment through Gazebo and MoveIt! Jointly, the UR5 robot arm is used to carry out the grasping sim-ulation task of four types of objects randomly placed on the table, and the average grasping success rate reaches 86%. The experimental results show that the improved algorithm has stronger adaptability to the grasping frame, and can effectively grasp both regular objects and objects with complex shapes and structures.
%K 抓取检测，注意力机制模块，ROS，Gazebo，MoveIt!<br>Grasp Detection
%K Attention Mechanism Module
%K ROS
%K Gazebo
%K MoveIt!
%U http://www.hanspub.org/journal/PaperInformation.aspx?PaperID=75169