哈尔滨工程大学学报2025，Vol.46Issue(10)：1942-1952,11.DOI:10.11990/jheu.202405007
基于姿态估计的物体运动三维测量方法
Three-dimensional measurement method of object motion based on pose estimation
摘要
Abstract
This study aims to address the limitations of traditional 3-D motion measurement methods,including high costs,low sampling rates,complex data fusion,and insufficient feature extraction.For this purpose,this work pro-poses a non-contact,non-destructive motion measurement method based on the DeepLabCut technology.Using clock hands as a model,we measured the 3-D motion trajectories of the keypoints by converting the high-dimension-al video signals of the clock keypoints into time-series datasets.After evaluating the training performance of 14 backbone networks,EfficientNet-B6 was chosen as the optimal backbone for high-precision line measurement,ena-bling the precise extraction of 2-D pixel coordinates of keypoints in binocular images.The 3-D coordinates were then calculated using the binocular camera calibration parameters and triangulation.The kinetic parameters of the keypoints were obtained by exploring the time-series datasets of the partial keypoint movements.Experimental re-sults reveal maximum relative errors of 2.49%for interkeypoint distance measurements and 2.27%for attitude an-gle measurements,with the visualizations accurately reconstructing clock hand movements.The proposed method presents strong adaptability and reliability in measuring the 3-D motion trajectories and kinematic parameters.Thus,it is a cost-effective and high-efficiency solution for the motion response monitoring of offshore structures.关键词
三维运动测量/非接触式/迁移学习/主干网络/双目相机标定/重投影误差/三角测量/坐标解算Key words
3-D motion measurement/non-contact/transfer learning/backbone network/binocular camera calibra-tion/reprojection error/triangulation/coordinate resolving分类
水利科学引用本文复制引用
霍家腾,刘鹏,王思琦,刘业宝,刘子洋,陈常龙..基于姿态估计的物体运动三维测量方法[J].哈尔滨工程大学学报,2025,46(10):1942-1952,11.基金项目
国家自然科学基金项目(52088102) （52088102）
山东省自然科学基金项目(ZR2022ME147). （ZR2022ME147）
