3D Object Model Fitting to Still Images Using Linear Combination Method of 2D Aspect Images

This paper describes a method for fitting 3D object model to still (single) 2D observed image by searching for the model's optimum posture parameters in the parameter space through LMM iteration method (a stabilized Gauss-Newton method). In the method, we propose a new approach to feeding an excellent initial parameter value to the iteration method via aspect identification by applying linear combination method of 2D aspect imaged[l], and via restricting the feature correspondence between the 2D image and the model's projection image to their circumferential features. Due to these schemes, the method gives a simple, fast and robust procedure. Numerical simulation using 500 randomly postured synthetic images of an irregular pentagonal prism shows 100% correct match only in 2.98 average iteration counts of the LMM method.