基于深度学习的增强装配智能人机交互方法
INTELLIGENT HUMAN-COMPUTER INTERACTION METHOD FOR AUGMENTED ASSEMBLY BASED ON DEEP LEARNING
李晗 1郭宇 1汤鹏洲 1吴鹏兴 1魏禛 1郑冠冠1
作者信息
- 1. 南京航空航天大学机电学院 江苏南京 210016
- 折叠
摘要
针对增强装配系统智能化水平低、系统与用户交互性差的问题,提出一种基于深度学习的零件拾取感知方法.采用目标检测算法提取手部和零件的类别与位置信息,并通过区域相交判定方法获取图像感兴趣区域.搭建分类网络以识别单帧深度图像中零件的拾取状态,以此为基础,提出连续帧状态识别方法消除单帧识别结果误差.实验结果表明,在航空发动机零件装配过程中,所提方法单帧感知速度在0.031 s内,连续帧下准确率可达100％,满足现场识别准确率和速度要求.
Abstract
Aimed at the problem of low level of intelligence of the augmented assembly system and poor interaction between the system and the user,a deep learning-based intelligent parts picking perception method is proposed.The target detection algorithm was used to extract the category and position information of the hand and the parts to be assembled,and the region intersection determination method was used to obtain the region-of-interest.A classification network was built to identify the pickup status of parts in a single frame depth image.Based on this,a continuous frame status recognition method was proposed to eliminate errors of single frame recognition result.The experimental results show that:in the assembly process of aero-engine parts,the proposed method has a single-frame sensing speed within 0.031 s,and the accuracy of continuous frames can reach 100％,which meets the accuracy and speed requirements of on-site recognition.
关键词
增强装配/人机交互/深度学习/拾取感知Key words
Augmented assembly/Human-computer interaction/Deep learning/Pickup perception引用本文复制引用
基金项目
国防基础科研资助项目(JCKY2018605C003)
国防基础科研资助项目(JCKY2018203A001)
国防基础科研资助项目(JCKY2017203B071)
出版年
2024
国家科技期刊平台
NSTL
维普
万方数据