Files

Abstract

Rooted in radio-astronomy, the Bluebild algorithm is a paradigm-changing imager to estimate the sky intensity function. However, current imaging pipelines in radio-astronomy are highly-tailored to the CLEAN-family of imagers, and hence lack the flexibility required to properly experiment with Bluebild and its associated technologies. Therefore, there is a need to design tools flexible enough to allow such algorithms to be implemented while retaining a large degree of freedom for future experimentation. Owing to its low computational complexity and affinity for parallel execution, Bluebild is a good candidate to achieve real-time imaging in the field. Moreover, its generality makes it broadly applicable to other domains as well, such as sound-field imaging. The goal of this thesis is therefore two-fold: First, we tackle the system aspects of Bluebild, working towards the goal of doing real-time imaging in radio-astronomy and acoustics. This is done by looking at the problem from an algorithm and architecture point-of-view. Secondly, we design Pypeline, a modular Python toolkit to solve the infrastructure crisis Bluebild faces at the moment. For radio astronomy, we leveraged algorithm speed-ups and system optimizations to reduce computation by an order of magnitude on CPU architectures, and showed that Bluebild can be efficiently implemented on GPUs by exploiting the inherent parallelism of the sampling stage. We additionally showed the first least-square image estimates using wide-angle beamformers. Based on the radio astronomy Bluebild implementation, we built a real-time sound field imager on top of the Pyramic microphone array that accounts for the specifics of acoustic arrays. We showed by experiment that reconstructions of Bluebild seem to be at least on par with those of B-scan and MUSIC, while being more stable than MVDR. Finally, we developed Pypeline, a Python package to design composable signal processing systems from elementary building blocks. Through its flexible design, it allowed us to implement entire imaging pipelines for many sensor arrays, allowing technologies such as Flexibeam and randomized beamforming to be tested for the first time with Bluebild.

Details

Actions

Preview