基于角度估计机械臂抓取系统目标检测算法  

Object Detection Algorithm of Manipulator Grasping System Based on Angle Estimation

作　　者：吕泽 蔡乐才 成奎 高祥 王新杰 LV Ze;CAI Lecai;CHENG Kui;GAO Xiang;WANG Xinjie(School of Automation and Information Engineerin,Sichuan University of Science&Engineering,Yibin 644000,China;School of Mechanical Engineering,Sichuan University of Science&Engineering,Yibin 644000,China;Sanjiang Research Institute of Artificial Intelligence and Robotics,Yibin University,Yibin 644000,China)

机构地区：[1]四川轻化工大学自动化与信息工程学院,四川宜宾644000 [2]四川轻化工大学机械工程学院,四川宜宾644000 [3]宜宾学院三江人工智能与机器人研究院,四川宜宾644000

出　　处：《四川轻化工大学学报（自然科学版）》2023年第2期46-56,共11页Journal of Sichuan University of Science & Engineering(Natural Science Edition)

基　　金：四川省科技厅面上项目(19ZDYF2284);宜宾市科技计划项目(2021GY008)。

摘　　要：针对传统工业机械臂只能对固定位置的固定目标进行抓取的问题,在YOLOv5的基础上提出了一种基于角度估计的机械臂抓取系统目标检测算法——MAR-YOLOv5。首先,为了满足机械臂抓取过程中对目标物体角度信息的需求,在原算法的基础上添加了角度估计模块,通过环形平滑标签(Circular Smooth Label,CSL)角度分类法精准预测目标物体角度。然后,通过增加目标检测层提高对小目标物体的检测能力。最后,在特征提取C3模块中加入了卷积注意力机制(Convolutional Block Attention Module,CBAM),以增强在复杂环境中特征提取的效果。实验结果表明:本算法在自制的增减材制造数据集上的检测速度达到48 FPS,检测精度(m AP)为92.68%,与只添加角度估计模块的YOLOv5算法相比,m AP提高了4.19%。对于目标密集、遮挡等情况,本算法也可实时、准确地完成检测任务。Aiming at the problem that traditional industrial manipulator can only grasp fixed targets at fixed positions,an object detection algorithm MAR-YOLOv5 based on angle estimation for manipulator grasping system has been proposed on the base of YOLOv5.Firstly,in order to meet the demand for the object angle information during the manipulator grasping process,the angle estimation module is added to the original model to accurately predict the object angle by Circular Smooth Label(CSL)angle classification.Secondly,the detection capability of small objects is improved by adding an object detection layer.Finally,a hybrid attention mechanism,Convolutional Block Attention Module(CBAM),is added to the feature extraction C3 module to enhance the effect of feature extraction in complex environments.The experimental results show that the proposed MAR-YOLOv5 algorithm achieves a detection speed of 48 FPS and a detection accuracy mAP of 92.68%on the homemade additive and subtractive manufacturing dataset,which is 4.19%higher than that of the YOLOv5 model with only addition of the angle estimation module.The model can also accomplish the detection task in real-time and accurately for the cases of dense objects and occlusion.

关 键 词：目标检测 机械臂抓取 角度估计 注意力机制 增减材制造 

分 类 号：TP241[自动化与计算机技术—检测技术与自动化装置]