摘要
位姿估计在工业场景进行零件的拣选抓取时扮演着非常重要的角色。但是,目前针对杂乱场景下工业弱纹理零件的6DoF(6 Degrees Of Freedom)位姿估计的研究还较少。特别是当这些工件使用相同的材质,且形状相近时,对位姿估计提出了更大的挑战。本文针对杂乱场景下工业弱纹理零件的6DoF位姿估计方法进行了研究。在杂乱场景下,首先使用了针对该场景下工业弱纹理零件位姿数据集的获取方法。接着,提出了一种基于PAVNet网络与注意力机制的学习框架。在该框架中,选用多阶段处理的方法从像素级对目标对象进行特征提取,再通过基于RANSAC的投票算法选出阈值范围内的关键点,最后通过求解这些关键点位置与工件的旋转和平移的关系,由此来估计其位姿。本文通过在公用数据集以及真实数据集上的实验验证了本文提出算法的精度,并且满足了工业应用要求。
Pose estimation plays a very important role in picking and grabbing parts in industrial scenes.However,there are few studies on 6 DoF(6 Degrees Of Freedom)pose estimation for industrial weakly textured parts in cluttered scenes.Especially when these workpieces use the same material and have similar shapes,it poses a greater challenge for pose estimation.This paper studies the 6 DoF pose estimation method for industrial weakly textured parts in cluttered scenes.In the cluttered scene,this paper uses the acquisition method for the pose dataset of industrial weakly textured parts in this scene,and proposes a learning framework based on PAVNet network and attention mechanism.In this framework,this paper first uses a multi-stage processing method to extract features in the target objects from the pixel level,then selects key points within the threshold range through a RANSAC-based voting algorithm,finally solves the relationship between the positions of these key points and rotation and translation of the workpiece to estimate its pose.Based on the above,this paper verifies the accuracy of the proposed algorithm through experiments on public datasets and real datasets,and meets the requirements of industrial applications.
作者
杨纯
陈权
王涛
YANG Chun;CHEN Quan;WANG Tao(School of Computer Science and Technology,Guangdong University of Technology,Guangzhou 510006,China)
出处
《智能计算机与应用》
2022年第6期13-19,27,共8页
Intelligent Computer and Applications
关键词
金属工件
弱纹理
位姿估计
像素级
metal workpiece
weakly texture
pose estimation
pixel-wise
