Topological maps model the environment as a graph, where nodes are distinctive places of the environment and edges indicate the relationships between them. They present an interesting alternative to the classic metric maps, due to their simplicity and storage needs, which convert them in an active research area. Several kinds of sensors have been used during years for topological mapping and localization. However, in the last decades, vision approaches have emerged because of the technology improvements and the amount of useful information that a camera can provide. In this paper, we review the main solutions presented in the last 15 years, and classify them in accordance to the kind of image descriptor employed. Advantages and disadvantages of each approach are thoroughly reviewed and discussed.