基于中心点对特征的无序堆叠工件位姿估计算法  

Pose estimation algorithm for disordered stacked workpieces based on central point pair features

作　　者：胡洪乐 吴国新[1] 左云波[1] 朱春梅[2] 杜俊波 周玲珑 HU Hongle;WU Guoxin;ZUO Yunbo;ZHU Chunmei;DU Junbo;ZHOU Linglong(Beijing Key Laboratory of Measurement and Control of Mechanical and Electrical System,Beijing Information Science&Technology University,Beijing 100192,China;Mechanical Electrical Engineering School,Beijing Information Science&Technology University,Beijing 100192,China)

机构地区：[1]北京信息科技大学机电系统测控北京市重点实验室,北京100192 [2]北京信息科技大学机电工程学院,北京100192

出　　处：《北京信息科技大学学报（自然科学版）》2024年第3期60-66,83,共8页Journal of Beijing Information Science and Technology University

基　　金：国家重点研发计划项目(2020YFB1713203)。

摘　　要：针对无序堆叠场景中工件相互遮挡导致的位姿估计精度下降的问题,提出一种基于中心点对特征的六维位姿估计算法。首先,模拟无序堆叠场景下点云目标多姿态随机摆放的物理环境,生成特征提取网络需要的带真实标签的数据集。进而,以中心点对特征构建离线阶段的全局特征描述。然后,对在线阶段的工件无序堆叠场景,利用动态图卷积神经网络(dynamic graph convolutional neural network, DGCNN)算法提取点云中心特征分数,确定工件中心点,并以此作为改进点对特征(point pair feature, PPF)算法的参考点。最后,使用IPA数据集和自采集场景数据对算法性能进行了验证。实验结果表明:所提算法降低了参考点选择的随机性,相比原始PPF算法在场景工件数目为30时平均准确率提升19.5百分点,5种工件场景下平均运行时间缩短29.00%。In response to the pose estimation accuracy degradation caused by mutual occlusion of workpieces in disordered stacking scenarios,a six-dimensional pose estimation algorithm based on central point pair features was proposed.Initially,a simulated physical environment was created to emulate disordered stacking scenarios,where point cloud targets were randomly positioned in multiple poses,generating a dataset with ground truth labels for the feature extraction network.Subsequently,an offline global feature description was constructed using central point pair features.Then,for the scenario of disordered stacked workpieces in the online stage,the dynamic graph convolutional neural network(DGCNN)algorithm was employed to extract central feature score from the point cloud,determining the feature scores for identifying the object′s central point.This central point was then utilized as a reference point for the enhancing point pair feature(PPF)algorithm.Finally,algorithm performance verification was conducted using the IPA dataset and self-collected scene data.Experimental results show that,the proposed algorithm reduced the randomness in reference point selection,resulting in an average accuracy improvement of 19.5 percentage points in scenarios with 30 workpieces compared with the original PPF algorithm.Moreover,the average runtime was reduced by approximately 29.00%across five different workpiece scenarios.

关 键 词：位姿估计 特征匹配 中心点提取 深度学习 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]