为了提高CGF(Learning Compact Geometric Features)的匹配精度,提出一种基于关键点曲率分类的特征提取方法.首先,计算点云的关键点并根据关键点的曲率将其分成七个类别.之后,针对不同的类别分别使用对该类别关键点具有鲁棒性的特征表征方法进行描述.最后,将所有的特征输入到设计的KPNN(Key Point Neural Network)神经网络中训练得到融合后的特征.其中,对于特征表征的方法使用参数可调的CGF特征描述子,通过实验确定不同参数组合的CGF分别用于表征不同类别的关键点.实验结果表明:在标准数据集上,相比于其他局部特征提取算法,该方法具有更高的匹配精度.
Abstract
In order to improve the matching accuracy of CGF(Learning Compact Geometric Features),a feature extraction method based on key point curvature classification is proposed.Firstly,the key points of the point cloud are calculated.Then,all the key points are divided into 7 categories according to the curvature of the key points.For different categories,a feature character-ization method that is robust to the key points of the category is used to describe it.Finally,all the features are input into the de-signed KPNN(Key Point Neural Network)for training to obtain the fused features.Among them,for the method of feature character-ization,a parameter-tunable CGF feature descriptor is used,and CGFs with different parameter combinations are determined through experiments to represent the different feature of key points on different categories.The experimental results show that our method has higher matching accuracy on different standard datasets while comparing with other state-of-art local feature extraction algorithms.
关键词
点云/特征提取/关键点/CGF/曲率
Key words
point cloud/feature extraction/key point/CGF/curvature
维普
万方数据