On-line Sparse Recovery from Correlated Multiple Measurements Without Time Latency

To eliminate the time latency between the observations and estimations during the on-line recovery for time-varying sparse signals from correlated multiple measurements, an on-line recovery algorithm without time latency is proposed in this paper. A sequential scheme is constructed for the support set detection and significantly higher detection accuracy is achieved compared with the existing algorithms. Expectation-maximization inference is introduced for the online estimation of the signals and model parameters, where the temporal correlation among signals is sufficiently exploited by a linear dynamic system and linearly optimal estimation for the signals is achieved through RTS smoothing in the E-stage of the inference process. Experiments showed that compared with existing algorithms, significantly higher reconstruction accuracy is obtained with acceptable reduction of efficiency.