基于RGB与深度信息融合的管片抓取位置测量方法  

Position measurement method for tunnel segment grabbing based on RGB and depth information fusion

作　　者：王林涛 毛齐 WANG Lin-tao;MAO Qi(School of Mechanical Engineering,Dalian University of Technology,Dalian 116024,China)

机构地区：[1]大连理工大学机械工程学院,辽宁大连116024

出　　处：《浙江大学学报（工学版）》2023年第1期47-54,共8页Journal of Zhejiang University：Engineering Science

摘　　要：针对管片拼装机在管片抓取阶段依赖人工的问题,提出用于自动化拼装的管片螺栓抓取阶段的位置测量方法.该方案通过SIFT算法匹配管片螺栓,利用Faster-Rcnn算法,筛选出位于工作区的待抓取的管片螺栓.通过添加注意力机制改变特征提取网络结构,使其在0.8的交并比下保持约94%的准确率并排除其他管片螺栓的识别干扰.在识别到目标管片螺栓后,结合深度相机的信息获取完整的三维坐标,使得抓取设备位于抓取位置时测量的管片螺栓位置各轴的误差均不超过3 mm,满足机械式抓取装置对抓取的精度要求.直接对管片螺栓进行识别,无须考虑管片在工作区摆放位置的不确定性造成的误差,避免了使用靶标进行测量时靶标与管片之间的相对位置误差及设置靶标的人力与时间成本.A position measurement method which fits automatic segment assembly to grab segment bolt was proposed in order to solve the problem that grabbing segments relied on manual labor. SIFT algorithm was used to match the target segment bolts, and Faster-Rcnn algorithm was used to choose the target segment placed in working area. Convolutional attention block module was implemented to change the structure of feature extracting network,which maintained the recognition accuracy around 94% under stricter IOU(intersection over union) of 0.8 and prevented the affect of other segment bolts. Information from depth camera was fused to get the complete three-dimensional coordination after target segment bolt being recognized. The error of measurement in each axis was less than 3 mm when grabbing facility was in position, which meeted the requirement of grabbing with mechanical hoisting facility. The segment bolt was directly recognized. Then the error caused by segment placement uncertainty need not be considered. The relative error between segment and target object and the manual and time cost of setting target object can be prevented.

关 键 词：SIFT特征匹配 Faster-Rcnn算法 注意力机制 信息融合 管片螺栓位置测量 

分 类 号：TP212[自动化与计算机技术—检测技术与自动化装置]