融合DeepLabV3神经网络的工件位姿检测研究  

Research of Workpiece Pose Estimation with DeepLabV3 Neural Network

作　　者：李嘉鑫 李天剑[1] 胡欢 黄民[1] LI Jia-xin;LI Tian-jian;HU Huan;HUANG Min(Mechanical and Electrical Engineering College,Beijing Information Science and Technology University,Beijing 100196,China)

机构地区：[1]北京信息科技大学机电工程学院,北京100196

出　　处：《机械设计与制造》2024年第7期331-335,共5页Machinery Design & Manufacture

基　　金：北京市科技计划项目—空地人立体化消防协同作战系统研发与示范应用(Z19110000141909)。

摘　　要：针对工业零件表面特征少无法使用基于特征的匹配算法以及使用传统基于模板的匹配算法在光照变化或背景混乱的场景下没有较好的鲁棒性。提出了一种基于DeepLabV3网络与传统的算法相结合的方法,极大的了提高了在复杂背景环境下的匹配结果和检测精度。首先在DeepLabV3网络下使用Hu矩和最小外接矩形法确定中心位置及旋转角度,第二步使用自适应的Harris角点检测与五点法相结合完成手眼相机的快速标定,最后在AUBO机械臂下完成定位实验,定位误差在0.5mm以内,结果表明该算法在复杂场景下和非均匀光照下有更好的表现。For industrial parts with few surface features,feature-based matching algorithms cannot be used,and traditional template-based matching algorithms are not robust in scenes which are varying lighting and have chaos background.So a method based on DeepLabV3 network that combine with traditional algorithms is proposed,which greatly improves the matching results and detection accuracy in complex background.First,with the Hu moment and the minimum bounding rectangle method we determine the center position and rotation angle under the DeepLabV3 network.In the second step,we use the adaptive Harris corner detection combined with the five-point method to complete the rapid calibration of the hand-eye camera,and finally complete experiment under the AUBO robotic arm.In the positioning experiment,the positioning error is within 0.5mm,and the results show that the algorithm has better performance in complex scenes and non-uniform illumination.

关 键 词：位姿检测 手眼系统 DeepLabV3 

分 类 号：TH16[机械工程—机械制造及自动化] TP242[自动化与计算机技术—检测技术与自动化装置]