A Unified Algorithmic Framework for Dynamic Compressive Sensing
Xiaozhi Liu, Yong Xia
TL;DR
The paper tackles recursive reconstruction of time-varying sparse signals from undersampled measurements by exploiting structured dynamic sparsity. It introduces the Partial-Laplacian sparsity model and a unified PLAY-CS framework that reveals connections among existing dynamic CS methods while generalizing to MMV via PLAY+ CS/MMV. An EM-based PLAY+ CS algorithm is proposed, with a broadband MMV extension that leverages dynamic joint sparsity. Across dynamic channel tracking and compressive background subtraction, the proposed methods outperform state-of-the-art approaches, demonstrating meaningful gains in practical, real-time wireless and video settings.
Abstract
We propose a unified dynamic tracking algorithmic framework (PLAY-CS) to reconstruct signal sequences with their intrinsic structured dynamic sparsity. By capitalizing on specific statistical assumptions concerning the dynamic filter of the signal sequences, the proposed framework exhibits versatility by encompassing various existing dynamic compressive sensing (DCS) algorithms. This is achieved through the incorporation of a newly proposed Partial-Laplacian filtering sparsity model, tailored to capture a more sophisticated dynamic sparsity. In practical scenarios such as dynamic channel tracking in wireless communications, the framework demonstrates enhanced performance compared to existing DCS algorithms.