Visual tracking using superpixel-based appearance model

Abstract

In this work, we propose a tracking algorithm that robustly handles complex variations in target appearance, scale, occlusion, and background. In particular, the algorithm exploits a novel superpixel-based appearance model for visual tracking. From the initial tracking window, we extract superpixels and compute their histogram features. In subsequent frames, we search for the region that maximizes the similarity of the superpixel features. Our algorithm detects target occlusion and updates the appearance model accordingly. As well, the model is updated to handle large-scale variations. We present experimental results on several publicly available challenging sequences. Qualitative and quantitative evaluation of our tracking algorithm show improved performance over state-of-the-art trackers. © 2013 Springer-Verlag.