Regresar

Kernel spectral clustering for motion tracking: A first approach

Abstract:

This work introduces a first approach to track moving-samples or frames matching each sample to a single meaningful value. This is done by combining the kernel spectral clustering with a feature relevance procedure that is extended to rank the frames in order to track the dynamic behavior along a frame sequence. We pose an optimization problem to determine the tracking vector, which is solved by the eigenvectors given by the clustering method. Unsupervised approaches are preferred since, for motion tracking applications, labeling is unavailable in practice. For experiments, two databases are considered: Motion Caption and an artificial three-moving Gaussian in which the mean changes per frame. Proposed clustering is compared with kernel K-means and Min-Cuts by using normalized mutual information and adjusted random index as metrics. Results are promising showing clearly that there exists a direct relationship between the proposed tracking vector and the dynamic behavior. © 2013 Springer-Verlag.