Parameter estimation and classification of machine parts based on specular or mirror-like image data (robot vision, shape from shading, surface classification)

Abstract

This thesis presents least squares estimation techniques for 3-D object location and orientation using the specular or mirror-like reflections seen from objects in a single image. The techniques are derived by calculating the exact equations specifying where specular reflections should be seen in images as a function of object type, location and orientation, and camera and light source location and orientation. Then the true parameters of the objects in the scene are estimated through suitable approximations. Equations are derived to approximate the conditions found in an average manufacturing environment as accurately as possible. The approach presented here may be used for estimating machine part location and orientation as an independent, self-contained scheme or in conjunction with other scene analysis operators. A specific type of concurrent operator may be one of many. A few examples of such operators are relaxation type surface region discriminators, texture segmenters and classifiers employing Markov Random Fields, constrained surface fitting for surface parameter estimation, or constrained line and curve fitting for estimation of object location and orientation. The estimates produced by the statistical operators of this thesis are fairly accurate, which is shown with some experiments with real image data; therefore, they are useful either as a first-stage approximator for object orientation and location or as a verification procedure for other approaches.