Using Dominant Sets for Object Tracking with Freely Moving Camera

Object tracking with freely moving cameras is an openissue, since background information cannot be exploited forforeground segmentation, and plain feature tracking is notrobust enough for target tracking, due to occlusions, distractors and object deformations. In order to deal withsuch challenging conditions a traditional approach, basedon Camshift-like color-based features, is augmented by introducing a structural model of the object to be tracked incorporating previous knowledge about the spatial relationsbetween the parts. Hence, an attributed graph is built ontop of the features extracted from each frame and a graphmatching technique is used to extract the optimal matchwith the model. Pixel-wise and object-wise comparisonwith other tracking techniques with respect to manually obtained ground truth are presented.