We propose a robust method for the identification of coronary arteries in computed tomography angiography (CTA) images. Utilizing geometric relations among the target and reference objects, which are assumed to follow a Gaussian distribution, an anatomic and geometric model is designed by Bayesian inference, which provides robust geometric priors for the target object localization. As a prerequisite process for the identification of coronary arteries, partially broken coronary artery segments found in CTA images are grouped and reconnected by geometric analysis of higher order curves connecting the broken segments. The geometric properties such as curvature and torsion represent naturalness and consistency between the vessel segments. As a problem to identify coronary arteries from CTA images, we demonstrate the robustness and accuracy of the proposed method in comparison with existing methods including commercial workstations on a variety of CTA cases.