A pure vision-based approach to topological SLAM