Incremental hybrid approach for unsupervised classification: Applications to visual landmarks recognition

Abstract

This paper describes a novel approach for incremental subspace learning which combines the best features of the evolving clustering method and the spectral clustering algorithm based on the graph p-Laplacian. The evolving clustering method is employed to classify each input sample into a set of spherically-shaped groups. Then, the spectral clustering algorithm is used to unsupervisedly cluster this reference set, resolving the shape of classes having non-zero covariance. The proposed approach has been applied to the problem of visual landmark recognition, in a mobile robot navigation framework. Experimental results show that the performance of the method is high in terms of error rate. © 2010 Springer-Verlag.