In this paper, we consider the problem of joint sparsity pattern recovery in a distributed sensor network. The sparse multiple measurement vector signals (MMVs) observed by all the nodes are assumed to have a common (but unknown) sparsity pattern. To accurately recover the common sparsity pattern in a decentralized manner with a low communication overhead of the network, we develop an algorithm named decentralized and collaborative subspace pursuit (DCSP). In DCSP, each node is required to perform three kinds of operations per iteration: 1) estimate the local sparsity pattern by finding the subspace that its measurement vector most probably lies in; 2) share its local sparsity pattern estimate with one-hop neighboring nodes; and 3) update the final sparsity pattern estimate by majority vote based fusion of all the local sparsity pattern estimates obtained in its neighborhood. The convergence of DCSP is proved and its communication overhead is quantitatively analyzed. We also propose another decentralized algorithm named generalized DCSP (GDCSP) by allowing more information exchange among neighboring nodes to further improve the accuracy of sparsity pattern recovery at the cost of increased communication overhead. Experimental results show that, 1) compared with existing decentralized algorithms, DCSP provides much better accuracy of sparsity pattern recovery at a comparable communication cost; and 2) the accuracy of GDCSP is very close to that of centralized processing.
- Compressive sensing
- Information fusion
- Joint sparsity pattern recovery
- Subspace pursuit
ASJC Scopus subject areas
- Signal Processing
- Electrical and Electronic Engineering