基于多尺度张量分解的点云结构特征提取

摘要

针对传统点云处理算法弱特征提取与抗噪声能力之间的矛盾,提出了一种基于多尺度张量分解的点云结构特征提取算法。首先,利用张量矩阵奇异值分解进行采样点特征显著性编码;然后,将法向（切向）一致性测度与罗曼诺夫斯基准则相结合求取采样点最优邻域,以提高采样点特征识别的可靠性;最后,利用最小生成森林进行特征点遍历,构建点云结构特征曲线。实验结果表明,该算法可以实现复杂点云结构特征的有效识别。%To solve the conflicts between the ability of weak feature extraction and noise resistency of traditional algorithms, a new feature extraction algorithm was proposed based on multi--scale ten- sor decomposition. Firstly, feature saliency encoding was defined based on the singular value decom- position of tensor matrix. Secondly, normal(tangential) consistent measure was constructed and used to determine the maxmuim scale combined with Romanovskii criterion. The reliability of the feature reconizing algorithm is improved. Finaly, the feature lines were constructed using minimal spanning forest. Expremental results reveal the weak feature extraction and noise resistency abilities of the method.