Abstract
Creating a 3D object model requires generally the acquisition of multiple partially overlapping range<br />images captured with different sensor locations. The 3D registration of these images that are assumed to<br />be pair-wisely overlapping consists in estimating the rigid transformation that places them in a common<br />coordinate system. This thesis presents solutions for automatically registering two unstructured 3D-color<br />point sets provided by a high-resolution scanner. Our first aim was to improve performance of classical<br />ICP (Iterative Closest Point) approaches by taking into account colour information for point matching.<br />When colour noise is not too important, using photometric information allows us to improve the<br />convergence and to decrease the residual error of the registration process. Moreover, it is possible to<br />reduce the drawbacks of intensity variations by choosing a perceptual colour space where intensity and<br />chromatic components are separated. The second aim of this work was to develop a fully automatic<br />registration method, which does not require any pose measuring hardware or manual intervention. In<br />fact, ICP requires an a-priori knowledge of an approximate estimate of the rigid transformation. This has<br />been solved using a set of interest point pairs extracted from the two colour images with the Harris<br />accurate detector. The initial 3D transform is then estimated by applying the RANSAC algorithm to the<br />set of 3D point pairs issued from the previous colour matching. This randomised estimator is robust to<br />outliers and provides a rough estimation that is enough accurate to initialise the iterative registration<br />process. Experimental results show the capabilities of the proposed solutions with colour range images<br />of real-world objects.