Model-based combined tracking and resolution enhancement

Abstract

Wide area surveillance requires high-resolution images of the object of interest derived possibly from only low-resolution video of the whole scene. We propose a combined tracking and resolution enhancement approach that increases the resolution of the object of interest during tracking. The key idea is the use of an object-specific 3D mesh model with which we are able to track non-planar objects across a large number of frames. This model is subdivided such that every triangle is smaller than a pixel when projected into the image to facilitate super-resolution on the model rather than on the image. We apply our approach to faces and show that it outperforms interpolation methods by achieving resolution enhancement, while being less blurred.