Registration of coronary arterial images taken at different times is very important for obtaining better visibility of differences between sequential images. A typical image registration algorithm often employs a similarity measure to detect the differences generated from the relative motion or gray level changes between these images. Although a number of image registration approaches have been proposed to resolve the registration problem of digital angiography, they are either computationally expensive or not very robust in the application to practical images. This paper presents a feature-based sum of absolute values of difference (SAVD) using a coarse-to-fine strategy. The proposed algorithm was demonstrated to be capable to automatically registering the arterial structures in the areas of interest selected from a pair of sequential images as well as providing fractional pixel precision in registration. Compared to other existing methods, the algorithm improves the speed and the reliability of registration when a pair of coronary arterial images are acquired at the same or almost the same phase of cardiac motion.