Distance metric for motion vector histograms based on human perceptual characteristics

This paper proposes a distance metric for motion vector histograms. Of a number of existing distance metrics for histograms, the weighted quadratic form distance is known to be efficient because of its flexibility. In order for the distance to be efficient, however, a suitable selection for its weighting matrix is necessary. Proposed in this paper is a determination technique for the weighting matrix which can reflect human perceptual characteristics for motion similarity. Experimental results have confirmed that the proposed technique is useful as a similarity measure in video retrieval.