Compressed sensing based image acquisition methodologies for constrained autonomous exploration systems with single pixel cameras

Kurzfassung

Scientific exploration of extraterrestrial worlds are carried out by robotic systems. Being far away from Earth, such robotic platforms require certain autonomy in their operation. They may be required to adapt to changing mission requirements or execute pre-planned mission objectives on their own. Moreover, they are constrained in the amount of electrical power available for various tasks. Thus fully autonomous scientific exploration requires acquisition strategies which can recognise informative regions in the instrument’s field-of-view and make on-the fly decisions of acquiring the regions with respect to resource constraints. This thesis addresses the problem of autonomous acquisition of Regions of Interest (RoIs) in an exploration scenario. The exploration scenario is modelled by a limited measurement budget constraint. Mechanical vibrations and stress of launching and landing the robotic platforms make moving parts susceptible to damage and malfunction. Such mechanical systems are more common for long wavelength (mid-IR, deep-IR, THz) imaging that use single pixel detectors or radiometers for image acquisition. The single pixel camera, which replaces the mechanical scanning radiometer with a Spatial Light Modulator (SLM), is studied as the imaging instrument in this thesis. An image acquisition methodology is developed for the single pixel imaging system with limited measurement budget that includes coarse sampling, region detection and segmentation, and compressed sensing. The basic acquisition algorithm arising from this methodology is named Measurement-Constrained Sampling (MCS). Compressed sensing is deployed in a imulti-level manner with random macro pixels with different macro pixel sizes and multi-level sampling of the Walsh transform. To introduce autonomy in the order of acquisition of the detected RoIs, two methods for RoI prioritisation are discussed. First, an empirical method based on the size of the detected RoI is considered. Second, a more nuanced method based on estimation of information change across spatial scales in the detected RoIs is developed. The estimator is evaluated for each RoI and is called the Refinement Indicator. The acquisition algorithm with the refinement indicator is named RoI Prioritised Measurement-Constrained Sampling (RP-MCS). The performance of the RP-MCS algorithm with multi-level sampling of the Walsh transform was found to be better than random macro pixel sampling at the considered measurement rates. A flexible approach to designing multi-level sampling patterns is proposed that enables better control over the choice of spatial frequency sub-bands and also the number of measurements in each of the sub-bands. To study the effect of the proposed acquisition methodology on the complete imaging pipeline, a plug-and-play framework for the instrument module of a remote exploration system is developed based on ESA guidelines. Influence of sampling parameters, noise, quantization and SLM damage on the complete imaging pipeline is studied and various trade-offs with respect to mission guidelines are discussed.