Topological localization based on key-frames selection and vocabulary tree for mobile robots

It is significant to realize topological localization for improving the ability of autonomous navigation for mobile robots in the unstructured environment. In this paper, we introduce vocabulary tree technique, which is popular and has been applied successfully in pattern recognition community, to the topological self-localization for mobile robots, and propose a topological localization algorithm based on key-frames selection and vocabulary tree. In the off-line phase for building robot´s topological map, we firstly propose a key-frames selection algorithm based on vocabulary tree, and pick up distinctive and discriminative keyframes. Then we extract local visual features from key-frames to build a vocabulary tree again, and obtain feature vectors which can sparsely represent local visual features of key-frames through vocabulary tree. These vectors are considered as the nodes of the topological map. In the on-line localization phase, we draw on a method of image retrieval based on vocabulary tree to realize robot self-localization. Finally, we use the COLD database to perform topological localization experiments, choose the local visual feature which is robust and can be run in real-time, and determine the best algorithm parameters. The effectiveness of the proposed algorithm is validated by the experimental results.