Coronary arteries registration can be used for coronary artery motion (CAM) quantification and even the prediction of cardiovascular disease. Accurate description of CAM can be analyzed from 3D Computer tomography angiography (CTA) image sequence stacks (which called 3D+t or 4D CTA images). In general, coronary arteries registration problems can be considered as a registration problem for the centerlines of coronary arteries, which can be settled by point sets registration approaches. However, the registration of coronary arteries in 4D CTA images remains a challenge mainly due to the irregular vascular shapes and non-rigid cardiac movement. To address such a challenge, highlighted by “Non-Rigid Point Set Registration by Preserving Global and Local Structures (PRGLS)”, an improved algorithm for registering coronary arteries with dynamic prior knowledge is proposed in this paper. This method can register coronary arteries in different phases with different shapes by capturing the global and local features of vessels with dynamic prior knowledge instead of the appointed prior knowledge in PRGLS. Experimental results demonstrate that proposed method is able to address the challenge of irregular vascular shapes in 4D cardiac CTA images and improve coronary arteries registration accuracy over the most existing methods.