Robust learning based 3D model registration and comparison -

Abstract

Nowadays and with the emerging technologies of 3D sensors-scanners and 3D printers, the problem of 3D model registration and comparison becomes a hot topic in the research community. 3D model registration is a challenging problem because the models obtained from two scans are registered in two unde ned coordinate systems. Similar to the problem of image registration or image stitching, 3D model registration looks for feature points in the subject models and tries to find the best fit based on these feature points. A method for 3D model registration and comparison is presented in this thesis. We make use of recent technologies, namely the Microsoft's Kinect, to obtain 3D reconstructed models. After obtaining the 3D models, we use point cloud registration techniques to compare 3D shapes and fit them together. Our system is able to detect any di erence between the two models being compared. Such a comparison mechanism is useful in localizing defects and-or damages in man-made products. For example, comparing a 3D scan of a dented car door to its un-dented counterpart helps localize the dent automatically and assesses the magnitude of the damage. The system can also be generalized to other applications such as model recovery problems, criminal investigations, archaeological preservation and quality checks.