An Efficient 3D Imaging using Structured Light Systems

Structured light 3D surface imaging has been crucial in the fields of image processing and computer vision, particularly in reconstruction, recognition and others. In this dissertation, we propose the approaches to development of an efficient 3D surface imaging system using structured light patterns including reconstruction, recognition and sampling criterion. To achieve an efficient reconstruction system, we address the problem in its many dimensions. In the first, we extract geometric 3D coordinates of an object which is illuminated by a set of concentric circular patterns and reflected to a 2D image plane. The relationship between the original and the deformed shape of the light patterns due to a surface shape provides sufficient 3D coordinates information. In the second, we consider system efficiency. The efficiency, which can be quantified by the size of data, is improved by reducing the number of circular patterns to be projected onto an object of interest. Akin to the Shannon-Nyquist Sampling Theorem, we derive the minimum number of circular patterns which sufficiently represents the target object with no considerable information loss. Specific geometric information (e.g. the highest curvature) of an object is key to deriving the minimum sampling density. In the third, the object, represented using the minimum number of patterns, has incomplete color information (i.e. color information is given a priori along with the curves). An interpolation is carried out to complete the photometric reconstruction. The results can be approximately reconstructed because the minimum number of the patterns may not exactly reconstruct the original object. But the result does not show considerable information loss, and the performance of an approximate reconstruction is evaluated by performing recognition or classification.;In an object recognition, we use facial curves which are deformed circular curves (patterns) on a target object. We simply carry out comparison between the facial curves of different faces or different expressions, and subsequently evaluate the performance of the reconstruction results. Since comparison between all pairs of curves can increase the computational complexity, we propose an approach for classification which is based on the shortest geodesic distances. Shape-based comparison is carried out because it shows robustness to scaling effect and rotation due to varying viewpoints. Previously, linear methods and non-linear methods have been investigated for a dimensional reduction which achieves efficient recognition / classification algorithms. But, existing approaches generate many parameters which leads to an optimization procedures which sometimes do not provide explicit solution. The proposed approach to dimensionality reduction for recognition is based on the property of the Fourier Transform whose magnitude response is symmetric and invariant to time-shift, and the results are much more explicit without loss of intrinsic information of targets.;In practice, to achieve the reconstruction of a larger sized object, we use multipleprojector-viewpoints (MPV) system. The minimum number of cameras and projectors is critical part to achieve an efficient MPV system. For an alternative view of reconstruction, we apply the concepts of a system identification to the reconstruction problem. The first one is a general system identification determined by the ratio of the output to input, and the second one is a modulation-demodulation theory used to estimate an input (transmitted) signal from an output (received or observed) signal.