Tensor Voting Based Foreground Object Extraction

Robust foreground extraction is necessary for good performance of any computer vision application such as tracking or video surveillance. In this paper, we propose a novel foreground extraction technique for static cameras which works for indoor as well as outdoor scenes. We model colors in a background frame by Gaussians using non-iterative tensor voting framework. For input frame, we compare color features of each pixel against background model and those that do not follow the model are classified as foreground pixels. We update background model to account for scene and lighting changes over time. In the case of significant background motion, we incorporate motion vectors within tensor voting framework to reduce misclassification. Experiments show that our approach is robust to background motion, noise, illumination fluctuations, scene and lighting changes.