TY - JOUR
T1 - 一种基于改进 DOPE 算法的螺栓位姿检测方法
AU - Wang, Xiangzhou
AU - Mei, Yunpeng
AU - Zheng, Shuhua
N1 - Publisher Copyright:
© 2023 Beijing Institute of Technology. All rights reserved.
PY - 2023/10
Y1 - 2023/10
N2 - To locate and acquire bolt pose accurately by the vision system fixed in the robot arm during the operation of the bolt fastening robot,a bolt pose detection study was conducted. Considering the rotational symmetry of bolt, a bolt pose detection method was proposed based on improved DOPE (deep object pose estimation) algorithm. Firstly, a backbone network with channel attention and parallel structure was designed to replace the VGG-19 backbone network in the DOPE algorithm, improving the detection capability of the model for small targets such as bolts. Secondly, setting the true value of the bolt pose during training as the best correspond to the prediction result among all potential symmetric poses of the target and proposing a loss function named Symloss, the improved detection method was arranged to solve the ambiguity problems caused by target symmetric pose possibly in detection process. Finally, improving scene modeling method and sampling method of pose data set, a NVISII ray-tracing engine was used to synthesize a high-quality anthropomorphic dataset. The experimental results show that in the experimental scenes of 30, 65 and 100 cm sight distance, the average distance pass rate of the model point for the bolt pose detection results is improved by 14.2%, 20.8% and 33.3% respectively compared with the DOPE algorithm under the threshold of 10% of the model diameter, and the detection speed is improved by 1.51 frames per second, improved the detection performance of the algorithm effectively.
AB - To locate and acquire bolt pose accurately by the vision system fixed in the robot arm during the operation of the bolt fastening robot,a bolt pose detection study was conducted. Considering the rotational symmetry of bolt, a bolt pose detection method was proposed based on improved DOPE (deep object pose estimation) algorithm. Firstly, a backbone network with channel attention and parallel structure was designed to replace the VGG-19 backbone network in the DOPE algorithm, improving the detection capability of the model for small targets such as bolts. Secondly, setting the true value of the bolt pose during training as the best correspond to the prediction result among all potential symmetric poses of the target and proposing a loss function named Symloss, the improved detection method was arranged to solve the ambiguity problems caused by target symmetric pose possibly in detection process. Finally, improving scene modeling method and sampling method of pose data set, a NVISII ray-tracing engine was used to synthesize a high-quality anthropomorphic dataset. The experimental results show that in the experimental scenes of 30, 65 and 100 cm sight distance, the average distance pass rate of the model point for the bolt pose detection results is improved by 14.2%, 20.8% and 33.3% respectively compared with the DOPE algorithm under the threshold of 10% of the model diameter, and the detection speed is improved by 1.51 frames per second, improved the detection performance of the algorithm effectively.
KW - data synthesis
KW - deep learning
KW - pose detection
UR - http://www.scopus.com/inward/record.url?scp=85177053312&partnerID=8YFLogxK
U2 - 10.15918/j.tbit1001-0645.2022.220
DO - 10.15918/j.tbit1001-0645.2022.220
M3 - 文章
AN - SCOPUS:85177053312
SN - 1001-0645
VL - 43
SP - 1094
EP - 1104
JO - Beijing Ligong Daxue Xuebao/Transaction of Beijing Institute of Technology
JF - Beijing Ligong Daxue Xuebao/Transaction of Beijing Institute of Technology
IS - 10
ER -