基于二维成像的三维物体形状特征分析

Shape features for projected 2D images based 3D object recognition

目前三维物体识别方法在识别过程中所需数据量大、难以实用。该文简化三维物体识别过程,构建了一个基于单视点二维投影图像的三维物体识别系统。分别选取Zernike矩、基于Trace变换的Triple特征、MSA等三种形状特征,实现了对物体的视点空间的聚类划分。在普林斯顿三维模型库上,通过分类识别实验分析三种形状特征的性能。实验表明:特征对不同类别物体的分类效果差异明显。该文由此提出了针对目标物体形状及应用环境的特征选取方案。

Most 3D object recognition methods need too much information for the recognition process so they are not practical for real applications. This paper presents a 3D recognition system based on a single projected 2D image to simplify the 3D recognition process. The system uses zernike moments, trace transformations, and multi-scale autoconvolution for the clustering based viewpoint space partitioning. A shape analysis of the Princeton shape benchmark is used to investigate the recognition of the three features on different 3D shapes. The results show that none of these features is suitable for all kinds of shapes. Therefore, a feature selection strategy is developed for recognition of different shapes.