Abstract:
Designing a discriminative speed function plays a vital role in conducting contour evolution in level set-based tracking framework. In this work, we propose a superpixel-driven speed function modeling method by fusion of two supplementary cues:appearance and motion. Based on kernel density estimation, a discriminative model separating the object from the background is constructed in appearance space. Meanwhile, by making use of the statistical characteristics of the optical flow field, the relative motion between the object and the background can be distinguished and enhanced by an adaptively chosen threshold. Finally, these two cues are combined in decision level under the Semi-Naive Bayes framework. Experimental results on a number of challenging video sequences demonstrate the effectiveness and robustness of the proposed tracking methods.