Real-time compressive sensing for small low power sensors: maximizing system performance in power impoverished environments

Small, light and low power sensors have been utilized in a growing number of modern applications. With rapid growth of circuit integration, we can fabricate more transistors onto a smaller area of silicon, providing sensors with increased computational capability. However, as the miniaturization continues, these sensors face constraints in size, weight and power: Despite rapid circuit integration, the size of the non-electronic components of the sensor such as batteries, optical and mechanical structures cannot be reduced following similar trend. Power consumption of the sensor must also be kept low to ensure long battery life and minimal heat dissipation. Due to these constraints, design for Size Weight and Power (SWaP) becomes a major challenge for miniature sensors. Compressed Sensing (CS), a signal processing theory, has shed light to SWaP constraints. CS theory suggests that the amount of samples a sensor must acquire to reconstruct a sparse signal can be much less than signal's Nyquist rate. Sparse signal is a signal that can be represented using only a few coefficients in a transform domain. In fact, in CS, sampling rate is independent from signal frequency, but is related to signal sparsity. This theory is important because many signals in nature exhibit sparsity: A pure tone sinusoidal is sparse in the frequency domain. A natural scene can also be represented sparsely using discrete cosine (DCT) basis. Acquiring these signal using less samples can greatly save system power consumption of the sensors. In this thesis, I demonstrate that CS can be used to mitigate the SWaP design constraints for miniature sensors. The analysis and discussion in this thesis focus on two specific design examples: In the first example, I discuss the SWaP constraints for high density neural recording device. I then demonstrate a real-time CS based neural recording system that achieves >90% communication power saving with efficient circuit area. In the second example, I first analyze the SWaP constraint and design trade-offs for miniature image sensors. I then demonstrate an all-CMOS implementation of pixel wise coded exposure imaging. This architecture can be used to implement spatiotemporal CS image sensor which reduces the system size, achieves power efficiency and improves image signal to noise ratio (SNR) and motion blur simultaneously. The same architecture can also be used to acquire high dynamic range (HDR) scenes using only single snapshot. Finally, I examine the common and different design considerations in these two examples at the conclusion chapter of the thesis