Region-Of-Interest Tracking

Short description

This tracking method relies on a statistical similarity measure in high dimension between a Region-Of-Interest (user-)defined in a keyframe (the reference ROI) and the regions potentially corresponding to it in subsequent frames. In a given frame, the region minimizing the similarity measure with respect to the reference ROI among admissible target regions is selected. Admissible regions are obtained by transforming the bounding box of the reference ROI according to a chosen motion model, e.g., translation and scaling.
The similarity measure is the Kullback-Leibler divergence between the multivariate feature PDF of the reference ROI and the one of a target region. Its estimation relies on k-th nearest neighbors. This allows to define features of high dimension combining radiometry (color, patches, gradient…) and geometry (see CVPR 2007 on the page Conferences and IEEE TIP 2009 on the page Journals for complete references and preprint).

Illustrative result

tracking result: frame 1 of sequence Crew tracking result: frame 11 of sequence Crew tracking result: frame 22 of sequence Crew tracking result: frame 31 of sequence Crew tracking result: frame 41 of sequence Crew tracking result: frame 51 of sequence Crew tracking result: frame 60 of sequence Crew tracking result: frame 72 of sequence Crew tracking result of sequence Crew: 3-D view