A two-stage approach to 3D scene structural analysis from a single two-dimensional image

Structural analysis of a 3D scene from 2D image pixel positions is an important and unresolved issue in image processing and pattern recognition. Current algorithms for 3D structural analysis do not explicitly use geometrical properties of distance between feature points and angle information as constraints to calibrate transformation parameters. We propose a novel two-stage approach to analyse the structure of a 3D scene based on the explicit expressions of distance and angle measurement. In the first stage, we transform the 3D-2D problem into a linear problem and propose a robust algorithm based on the total least squares method to estimate the rotation matrix and translation vector. In the second stage, we propose a coarse-to-fine geometric algorithm to refine the estimation. The proposed two-stage approach provides the closed form solutions to structural parameters through the analysis of image correspondence vectors synthesised into a single coordinate frame. For a comparative study of performance, we also extended a well known 2D-2D epipolar geometry algorithm proposed by R.Y. Tsai and T.S. Huang (1984) to solve 3D-2D problems. Experimental results show that the proposed approach and geometric algorithms are in general more accurate, stable, and efficient than the extended Tsai and Xuang algorithm.