Denoising-Based Turbo Message Passing for Compressed Video Background Subtraction

In this paper, we consider the compressed video background subtraction problem that separates and foreground of a from its measurements. The usually lies in low dimensional space is sparse. More importantly, each frame natural image has textural patterns. By exploiting these properties, develop message passing algorithm termed offline denoising-based turbo (DTMP). We show structural properties can be efficiently handled by existing denoising techniques under framework. further extend DTMP to online scenario where data collected an manner. extension based on similarity/continuity between adjacent frames. adopt optical flow method refine estimation foreground. also sliding window reduce complexity. Gaussianity messages, state evolution characterize per-iteration performance DTMP. Comparing algorithms, work at much lower compression rates, subtract successfully with mean squared error better visual quality for both subtraction.