The problem of appearance-based mapping and navigation in outdoor environments is far from trivial. In this paper, an appearance-based topological map, covering a large, mixed indoor and outdoor environment, is built incrementally by using panoramic images. The map is based on image similarity, so that the resulting segmentation of the world corresponds closely to the human concept of a place. Using high-resolution images and the epipolar constraint, the resulting map is shown to be very suitable for localization, even when the environment has undergone seasonal changes.

This paper presents a novel use of spectral clustering algorithms to support cases where the entries in the affinity matrix are costly to compute. The method is incremental - the spectral clustering algorithm is applied to the affinity matrix after each row/column is added - which makes it possible to inspect the clusters as new data points are added. The method is well suited to the problem of appearance-based, on-line topological mapping for mobile robots. In this problem domain, we show that we can reduce environment-dependent parameters of the clustering algorithm to just a single, intuitive parameter. Experimental results in large outdoor and indoor environments show that we can close loops correctly by computing only a fraction of the entries in the affinity matrix. The accompanying video clip shows how an example map is produced by the algorithm.

This paper presents an algorithm that builds topological maps, using omnidirectional vision as the only sensor modality. Local features are extracted from images obtained in sequence, and are used both to cluster the images into nodes and to detect links between the nodes. The algorithm is incremental, reducing the computational requirements of the corresponding batch algorithm. Experimental results in a complex, indoor environment show that the algorithm produces topologically correct maps, closing loops without suffering from perceptual aliasing or false links. Robustness to lighting variations was further demonstrated by building correct maps from combined multiple datasets collected over a period of 2 months