基于Top-down范式的施工现场钢结构焊缝坡口特征实时检测方法  

Real-time detection method for weld feature of construction site steel structure based on top-down paradigm

作　　者：成佳明 靳慧[1,2] 郑子健 蒋朗坤 罗琴丽 董凯 周军红 陈小飞 Cheng Jiaming;Jin Hui;Zheng Zijian;Jiang Langkun;Luo Qinli;Dong Kai;Zhou Junhong;Chen Xiaofei(School of Civil Engineering,Southeast University,Nanjing 211189,China;Jiangsu Key Laboratory of Engineering Mechanics,Southeast University,Nanjing 210096,China;China Construction Science and Industry Jiangsu Corporation Ltd.,Nanjing 211106,China;China Construction Steel Structure Jiangsu Corporation Ltd.,Taizhou 214532,China)

机构地区：[1]东南大学土木工程学院,南京211189 [2]东南大学江苏省工程力学分析重点实验室,南京210096 [3]中建科工集团江苏有限公司,南京211106 [4]中建钢构江苏有限公司,泰州214532

出　　处：《东南大学学报（自然科学版）》2023年第6期1100-1110,共11页Journal of Southeast University：Natural Science Edition

基　　金：国家自然科学基金资助项目(51578137)。

摘　　要：针对现场焊缝图像背景复杂、坡口特征点定位难度大的问题,提出了一种端对端的焊缝特征实时检测方法,用于提升建筑钢结构焊接效率和质量.基于人体姿态估计的思想,将焊缝特征点提取等效于人体骨骼关键点检测任务,遵从姿态估计Top-down范式建立施工现场建筑钢结构焊缝坡口特征提取方法.首先引入RTMdet目标检测器快速定位焊缝坡口区域,随后基于RTMPose姿态估计模型对目标区域进行坡口特征点检测,该网络将特征点坐标回归定位转换为特征点横纵坐标分类问题,有效提升了特征点定位的精度和效率.实验结果表明,相比基于数字图像处理的焊缝识别方法和基于全连接层回归的焊缝识别方法,该方法能够在包含复杂信息的焊缝图像中快速准确地提取焊缝特征点,单幅图像的特征点定位误差小于2像素,平均处理时间为38.2 ms,能满足施工现场自动化焊接的需求.Aiming at the problems of complex background textures in on-site weld images and difficulty in locating profile feature points,a real-time detection method for end-to-end weld features is proposed to improve the welding efficiency and quality of building steel structures.Adopting the idea from human pose estimation,the extraction of weld feature points is equivalent to the key point detection task of the human skeleton.Following the top-down paradigm of pose estimation,the extraction method for weld feature of building steel structure in construction site is established.The RTMdet object detector is first introduced to quickly locate the welding profile area,then the RTMPose pose estimation model is applied to detect profile feature points in the target region.By converting the feature point coordinate regression localization into a classification problem of horizontal and vertical coordinates,the localization accuracy and efficiency is effectively improved.Experimental results demonstrate that compared to the welding identification methods based on digital image processing or regression with fully connected layers,the method can rapidly and precisely extract the welding feature points from images containing complex information.The localization error of the feature point in a single image is less than 2 pixels,and the average processing time is 38.2 ms,meeting the requirements for automated welding on construction sites.

关 键 词：智能建造 结构光视觉 深度学习 人体姿态估计 焊缝特征点 

分 类 号：TU74[建筑科学—建筑技术科学]