Computer Engineering and Applications ›› 2023, Vol. 59 ›› Issue (23): 175-182.DOI: 10.3778/j.issn.1002-8331.2305-0140

• Graphics and Image Processing • Previous Articles     Next Articles

Research on DPC Clustering Multi-Objective Detection Method for Disorderly Grasping

CHEN Zeyu, LI Xiangguo, CAO Dengfeng, ZHU Denglin   

  1. College of Mechanical and Electrical Engineering,Hohai University, Changzhou,Jiangsu 213022, China
  • Online:2023-12-01 Published:2023-12-01

面向无序抓取的DPC聚类多目标检测方法研究

陈泽瑜,李向国,曹登锋,朱灯林   

  1. 河海大学 机电工程学院,江苏 常州 213022

Abstract: In order to detect as many graspable targets as possible from the scene, a multi-target detection algorithm based on DPC feature point clustering is proposed. The feature points of the template image and the image to be detected are first extracted using the SIFT algorithm, and clustered using the DPC algorithm to obtain the set of feature points belonging to different clustering centres. Then the feature points belonging to different clustering centres are matched with the template image feature points, combined with the RANSAC algorithm to remove false matches and count the number of correct matches, and the homography matrix from the template image to the image to be detected is calculated based on the correctly matched feature points to obtain the target detection result. Finally, the correct detection results are filtered according to the number of correct matching points for each target, and among the correct detection results, the graspable targets are filtered according to the difference between the number of matching points and the maximum number of matching points for the target. After detecting the graspable target, the scene disparity map is obtained using the stereo matching algorithm, the 3D coordinates of the target are calculated, and the target pose is calculated using the PNP algorithm based on the correspondence between the 3D coordinates and the 2D coordinates in the image. The experimental results show that the DPC clustering-based multi-target detection method can accurately detect the target object among multiple identical stacked targets and calculate the target pose separately, effectively solving the problem of multi-target detection in unordered grasping applications.

Key words: disordered grasping, template matching, density peak clustering(DPC), stack target

摘要: 为了尽可能多地从场景中检测出可抓取目标,提出了一种基于DPC特征点聚类的多目标检测算法。使用SIFT算法提取模板图像和待检测图像的特征点,并使用DPC算法对待检测图像特征点聚类,得到属于不同聚类中心的特征点集合。将属于不同聚类中心的特征点分别与模板图像特征点进行匹配,结合RANSAC算法去除误匹配并统计正确匹配点数量,根据正确匹配的特征点计算从模板图像到待检测图像的单应矩阵从而得到目标检测结果。根据每个目标正确匹配点数量筛选正确的检测结果,并在正确的检测结果中根据目标匹配点数量和目标最多匹配点数量的差值筛选出可抓取目标。检测出可抓取目标之后,使用立体匹配算法得到场景视差图,计算目标的三维坐标,并根据三维坐标与图像中二维坐标的对应关系使用PNP算法计算目标位姿。实验结果表明,基于DPC聚类的多目标检测方法能够在多个相同堆叠目标中准确检测出目标物体并分别计算位姿,有效解决了无序抓取应用中的多目标检测问题。

关键词: 无序抓取, 模板匹配, 密度峰值聚类(DPC), 堆叠目标