Internal shape-deformation invariant 3D surface matching using 2D principal component analysis

摘要

This paper describes a method that overcomes the problem of internal deformations in three-dimensional (3D) range image identification. Internal deformations can be caused by several factors including stereo camera-pair misalignment, surface irregularities, active vision methods' incompatibilities, image imperfections, and changes in illumination sources. Most 3D surface matching systems suffer from these changes and their performances are significantly degraded unless deformations' effect is compensated. Here, we propose an internal compensation method based on the two-dimensional (2D) principal component analysis (PCA). The depth map of a 3D range image is first thresholded using Otsu's optimal threshold selection criterion to discard the background information. The detected volumetric shape is normalized in the spatial plane and aligned with a reference coordinate system for rotation-, translation- and scaling-invariant classification. The preprocessed range image is then divided into 16x16 sub-blocks, each of which is smoothed to minimize the local variations. The 2DPCA is applied to the resultant range data and the corresponding principal vectors are used as the characteristic features of the object to determine its identity in the database of pre-recorded shapes. The system's performance is tested against the several 3D facial images possessing arbitrary deformation. Experiments have resulted in 92% recognition accuracy for the GavaDB 3D-face database entries and their Gaussian- or Poisson-type noisy versions using the minimum Euclidean-distance classification strategy in an optimally constructed eigen-face feature space.