Kernel-based unsupervised trajectory clusters discovery

Nowadays support vector machines (SVM) are among the most popular tools for data clustering. Even though the basic SVM technique works only for 2-classes problems, in the last years many variants of the original approach have been proposed, such as multi-class SVM for multiple class problems and single-class SVM for outlier detection. However, the former is based on a supervised approach, and the number of classes must be known a-priori; the latter performs unsupervised learning, but it can only discriminate between normal and outlier data. In this paper we propose a novel technique for data clustering when the number of classes is unknown. The proposed approach is inspired by single-class SVM theory and exploits some geometrical properties of the feature space of Gaussian kernels. Experimental results are given with special focus on the field of trajectory clustering1.