基于曲率关键点的点对特征三维目标识别
Three-dimensional object recognition based on point-pair features of curvature key points
邓天睿 1刘冉 1肖宇峰 1郭林 1蓝发籍 1王林1
作者信息
- 1. 西南科技大学信息工程学院,四川绵阳 621000;特殊环境机器人技术四川省重点实验室,四川绵阳 621000
- 折叠
摘要
精准的三维(three-dimensional,3D)目标识别对于机器人自主抓取至关重要,针对目前基于原始点对特征(point-pair feature,PPF)的三维目标识别算法中存在识别速度慢、严重遮挡场景下识别率低的问题,提出了一种基于曲率关键点的点对特征三维目标识别算法.该算法根据点云法向量邻域夹角均值,快速估算点云曲率,以此提取关键点,通过对关键点计算点对特征,剔除了模型点对特征哈希表中存在的大量冗余点对.使用结合位姿聚类和假设检验的位姿优化算法,首先通过位姿聚类对候选假设位姿进行优化,其次位姿聚类后采用ICP(iterative closest point)算法对候选位姿进行细化,最后利用基于重合度计算匹配分数的假设检验算法滤除错误假设并得出最佳假设位姿.实验结果表明,在公开数据集上,所提方法能够获得95.2％的平均识别率,减少模型点对特征哈希表构建时间并且提高在严重遮挡场景下的识别率.
Abstract
Three-dimensional(3D)target recognition plays a critical role in autonomous robot grasping.The original point-pair feature-based 3D target recognition methods are facing the challenges,such as slow recognition speed and poor recognition accuracy in severely occluded scenes.To address these is-sues,this paper proposes a 3D target recognition method based on the point-pair feature of curvature key points.First,the key points are extracted based on the curvature,which is estimated by mean angle of neighbored point cloud normal vector in a point cloud.Second,the point-pair features are established for key points to eliminate redundant point pairs in the point-pair feature Hash table.Third,through proces-ses of pose clustering and hypothesis verification,the pose optimization is achieved.In particular,we opti-mize the candidate poses by pose clustering and then refine the pose by the iterative closest point(ICP)algorithm.Finally,the incorrect poses are eliminated by the hypothesis verification of poses according to the overlap score.The hypothesis of optimal posture is obtained.The experiment results show that the proposed method achieves an average recognition accuracy of 95.2％on a public dataset,and significantly reduce the construction time of the point-pair feature Hash table,while enhancing the recognition per-formance in severely occluded scenes.
关键词
三维(3D)目标识别/点对特征/关键点提取/假设检验Key words
three-dimensional(3D)object recognition/point-pair feature/key point extraction/hypothe-sis verification引用本文复制引用
基金项目
国家自然科学基金(12175187)
国家自然科学基金(12205245)
国家重点研发计划(2019YFB1310805)
四川省自然科学基金(2023NSFSC0505)
出版年
2024
NETL
NSTL
万方数据