Compressive sensing for computer vision

 

We propose a method to directly recover background-subtracted images using compressive sensing (CS), which is attractive for communication constrained multi-camera computer vision problems. We apply the CS theory to recover object silhouettes (binary background subtracted images) when the objects of interest occupy a small portion of the camera view, i.e., when they are sparse in the spatial domain. We cast the background subtraction as a sparse approximation problem and provide different solutions based on convex optimization and total variation. In our CS-based method, as opposed to learning the background, we learn and adapt a low dimensional compressed representation of it, which is sufficient to determine spatial innovations; object silhouettes are then estimated directly using the CS measurements without any auxiliary image reconstruction. Our approach is suitable for image coding in communication constrained problems by using data captured by multiple conventional cameras to provide 2D tracking and 3D shape reconstruction results with compressive measurements.

 

Authors: Richard G. Baraniuk, Volkan Cevher, Marco F. Duarte

Publications: Preprint, NIPS 2008

Related research at Rice