The introduction of compressive sensing (CS) led to a new paradigm in signal processing. Traditionally, signals are sampled above their Nyquist rate. Using CS, the same information is acquired with much fewer measurements, provided a sparse representation of the signal exists. This makes CS a very promising technology with a large number of potential applications. While the acquisition of measurements is simplified, the reconstruction of the original signal becomes more involved. Sparse signal recovery algorithms solve the corresponding systems of under-determined linear equations and have proven very efficient for various applications. Examples include de-noising, the restoration of corrupted signals, signal separation, super-resolution, and in-painting. All applications are based on the observation that many natural and man-made signals have sparse representations in some suitable bases. In the last few years, impressive progress has been made in the development and characterization of fast recovery algorithms. However, the computational effort for successful signal recovery remains high, even for problems of moderate size. Reconstruction becomes especially challenging for real-time applications with stringent power constraints, e.g., on mobile devices. Such applications require efficient hardware implementations of sparse signal recovery algorithms, which we develop in this thesis. We present different architectures of greedy algorithms for a number of selected applications. The first example is the estimation of sparse channels in broadband wireless communication. The use of sparse recovery algorithms efficiently reduces noise and, thus, increases estimation quality. Architectures for three algorithms are developed and their realizations in ASICs are compared. We show that approximative algorithms deliver good results at low hardware complexity. The second application is the recovery of signals corrupted by structured noise. Using the example of audio restoration from corruptions by clicks and pops, fast realizations of the approximate message passing algorithm are designed. Two fundamentally different architectures -one relying on fast transforms, the other relying on parallel processing power- are developed and compared. Large gains in terms of throughput and circuit complexity are realized by applying fast transforms in the context of CS recovery. The choice of the most attractive algorithms and architectures depends on the sparsity, the number of measurements, and the basis in which the samples are taken. In general, approximate message passing is found to be very well suited for hardware recovery of moderately sparse signals while serial greedy pursuits are better suited for very sparse signals. Further, a new application of CS in localization is explored. We show how sparse recovery increases the detection accuracy in passive radar systems based on WiFi signals. Finally, also a new sensing device, acquiring measurements with very low hardware complexity, is introduced. This modified analog-to-digital converter samples at non-uniformly distributed points, which allows the reconstruction of Fourier-sparse signals from very few measurements. All the presented examples and hardware implementations bring CS one step closer to practical applications. Patrick Mächler was born in Zurich, Switzerland, in 1984. He received his MSc degree in information technology and electrical engineering from ETH Zurich, Switzerland, in 2008. In 2008, he was a visiting researcher at Berkeley Wireless Research Center (BWRC), UC Berkeley, CA, USA. In the same year, he joined the Integrated Systems Laboratory of ETH Zurich as a research assistant. His research interests include digital signal processing, VLSI architectures, compressive sensing, and wireless communication.weiterlesen