Segmentation of blood vessels in retinal images allows early diagnosis of disease; automating this process images is an important and challenging task in medical analysis and diagnosis. The main goal of our work is an extraction of the vessel centerline. The method is based on a tracking strategy, the vessel is assumed to be represented by a succession of spatially oriented cylinders, and its axis is obtained by the concatenation of the principal axes of inertia of the constituent cylinders. The centerline extraction is done by an iterative prediction estimation tracking technique based on a multi-scale analysis of image moments and on a shape model close to snakes. Experimental results are presented in Retinal images.