摘要
针对现有的抓取位姿检测算法难以实现高精度按类抓取的问题,设计了一种基于深度和语义信息的抓取位姿检测网络YLG-CNN。首先,对抓取检测算法GG-CNN进行改进,在特征提取模块添加残差结构,以融合不同层次的特征,提升检测模型对深度信息的理解能力,在残差结构末端引入CBAM注意力机制,使宜于作为抓取中心的热力像素点获得更高的热力值,得到更为可靠的抓取位姿;其次,通过YOLOv5算法获取待抓取目标的类别,并将其映射到改进GG-CNN所输出的抓取热力图中,为每个抓取点赋予抓取对象的语义信息,实现按类抓取。最后,设计了一套基于机器人操作系统(ROS)的3D视觉智能抓取系统,通过真实抓取实验验证所提方法。实验结果表明,所提的残差注意力抓取网络可抓取精度达到78.2%,较次优算法GGCNN+CBAM提高6.8%,并且YLG-CNN分类抓取网络能够实现高精度的分类抓取,其平均抓取成功率达到78.3%,较于GG-CNN+YOLOv5算法提升了17.1%。
In order to solve the problem that the existing grasping pose detection algorithms are difficult to achieve high-precision of class based grasping,a grasping pose detection network YLG-CNN based on depth and semantic information has been proposed.Firstly,the grasping detection algorithm GG-CNN is improved by adding residual structures in the feature extraction module to fuse different levels of features and enhance the detection model to understand the depth information.CBAM attention mechanism is introduced into the residual structure to obtain higher thermal values for thermal pixels that are suitable for grasping,resulting in more reliable grasping poses.Secondly,the YOLOv5 algorithm is used to obtain the category of the target to be captured,and map it to the capture heat map output by the improved GG-CNN,giving each capture point the semantic information of the target to be captured,so as to achieve the capture by category.Finally,a 3D visual intelligent grasping system based on robot operating system(ROS)is designed,and the proposed method is verified through real grasping experiments.The experimental results show that the proposed residual attention grasping network can achieve a grasping accuracy of 78.2%,which is 6.8%higher than the GG-CNN+CBAM algorithm.Moreover,the YLG-CNN classification grasping network can achieve high-precision classification grabbing,with an average grasping success rate of 78.3%,which has increased by 17.1%compared with the GG-CNN+YOLOv5 algorithm.
作者
王艺成
张国良
汪坤
张自杰
WANG Yicheng;ZHANG Guoliang;WANG Kun;ZHANG Zijie(School of Automation and Information Engineering,Sichuan University of Science&Engineering,Yibin 644000,China;Artificial Intelligence Key Laboratory of Sichuan Province,Yibin 644000,China)
出处
《四川轻化工大学学报(自然科学版)》
CAS
2024年第5期78-86,共9页
Journal of Sichuan University of Science & Engineering(Natural Science Edition)
基金
四川省科技厅重点研发项目(2023YFG0196)。
关键词
目标识别
抓取位姿检测
GG-CNN
残差网络
分类抓取
object recognition
grasping pose detection
GG-CNN
residual network
classification grasping