Fast implementation of sparse reconstruction for CS-based DoA estimation

Abstract

Sparse vector reconstruction requires a long computation time, because it is based on some iterative computation algorithms, in which an initial dense vector is gradually modified to a sparse vector. To overcome this problem, we proposed a fast implementation technique that is based on the reordering/reuse of results calculated from the zero-elements at each iteration. In addition, we adapted our technique to a GPU (graphics processing unit)-suitable implementation of ℓp-norm minimization, i.e., a CS (compressive/compressed sensing)-based DoA (direction of arrival) estimation algorithm. We found that the proposed implementation with a GPU is up to 47 times faster than the conventional implementation with an 8-threaded CPU.