一种基于改进DOPE算法的螺栓位姿检测方法  

A Bolt Pose Detection Method Based on Improved DOPE Algorithm

作　　者：王向周[1] 梅云鹏 郑戍华[1] WANG Xiangzhou;MEI Yunpeng;ZHENG Shuhua(School of Automation,Beijing Institute of Technology,Beijing 100081,China)

机构地区：[1]北京理工大学自动化学院,北京100081

出　　处：《北京理工大学学报》2023年第10期1094-1104,共11页Transactions of Beijing Institute of Technology

基　　金：国家电网有限公司总部科技项目(5200-2020036147A-0-0-00)。

摘　　要：针对螺栓紧固机器人在作业过程中机械臂所载视觉系统定位并获取螺栓姿态的任务,进行了螺栓位姿检测研究.针对螺栓具有旋转对称性的特点,提出一种基于改进深度物体姿态估计(deep object pose estimation,DOPE)算法的螺栓位姿检测方法.首先,设计了一种具有通道注意力及并行结构的骨干网络替换DOPE算法中的VGG-19骨干网络,提高了模型对于螺栓等小型目标的检测能力;其次,将训练过程中螺栓姿态的真实值设定为预测结果在螺栓所有潜在对称姿态中的最优匹配,并据此提出Symloss损失函数,解决了目标对称姿态可能在检测过程中引起歧义的问题;最后,对姿态数据集的场景建模方法与采样方式进行改进,利用NVISII光线追踪引擎合成高质量拟真数据集.实验结果显示,在30、65和100 cm三种视距的实验场景中,本算法对螺栓姿态检测结果的模型点平均距离通过率在模型直径10%阈值下较DOPE算法分别提高14.2%、20.8%与33.3%,检测速度提高1.51帧/s,有效提高了算法对于螺栓姿态的检测性能.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.

关 键 词：深度学习 姿态检测 数据合成 

分 类 号：TP249[自动化与计算机技术—检测技术与自动化装置]