Adaptive optics imaging of the retina has recently proven its capability to image micrometric structures such as blood vessels, involved in common ocular diseases. In this paper, we propose an approach for automatically segmenting the walls of retinal arteries in the images acquired with this technology. The walls are modeled as four curves approximately parallel to a previously detected reference line located near the vessel center (axial reflection). These curves are first initialized using a tracking procedure and then more accurately positioned using an active contour model embedding a parallelism constraint. We consider both healthy and pathological subjects in the same framework and show that the proposed method applies in all cases. Extensive experiments are also proposed, by analyzing the robustness of the axial reflections detection, the influence of the tracking parameters as well as the performance of the tracking and the active contour model. Noticeably, the results show a good robustness for detecting axial reflections and a moderate influence of the tracking parameters. Compared to a naive initialization, the active contour model coupled with the tracking also offers faster convergence and better accuracy while keeping an overall error smaller or very near the inter-physicians error.