Abstract
This dissertation presents the design and implementation of hardware-software co-integrated architectures that enable real-time, energy-efficient sensing at the edge. To overcome the limitations of conventional centralized computing, particularly in latency, data transfer, and power efficiency, this work investigates intelligent edge systems that perform computation in close proximity to the sensor.
The proposed approach integrates in-sensor and near-sensor computing to enable direct signal encoding, analog-domain processing, and localized inference. In-sensor computing is achieved through the co-location of sensing and memory elements, allowing real-time signal modulation and data reduction at the pixel level. Near-sensor computing complements this by executing neural network-based processing adjacent to the sensor, supporting fast and efficient interpretation of encoded information without reliance on external systems. These computing paradigms are further enhanced through systematic optimization of hardware and software components to minimize computational complexity and energy usage to fit in miniaturized system, while maintaining task accuracy and responsiveness.
Building on these capabilities, a unified sensing framework is introduced, drawing inspiration from biological vision. By coordinating wide-area awareness and focused high-resolution imaging, the system adapts dynamically to environmental stimuli while maintaining low power consumption and minimal data movement. Altogether, this dissertation establishes a foundational platform for intelligent sensing, demonstrating how edge-integrated architectures can advance the performance, efficiency, and autonomy of next-generation sensing systems.
