MIXED-SIGNAL IN-SENSOR COMPUTING AND COMMUNICATION TECHNIQUES FOR ULTRA-LOW-POWER VIDEO SENSOR NODES

Abstract

Video Sensor Nodes (VSNs) for Internet-of-Things (IoT) and Internet-of-Bodies (IoB) applications operate under severe energy and form-factor constraints imposed by continuous sensing, high data rates, and limited energy availability. These challenges become particularly critical in wearable and implantable systems, where miniature batteries restrict operating lifetime and frequent battery replacement may be impractical. At the same time, continuously acquired visual information introduces significant energy overhead from analog-to-digital conversion, computation, storage, and communication. This dissertation addresses these challenges through mixed-signal and time-domain techniques that reduce energy across the sensing--computation--communication chain, together with wireless power delivery for energy-constrained biomedical systems. The first part of this dissertation addresses the energy associated with front-end digitization and communication. Conventional Video Sensor Nodes typically digitize analog pixel values using an ADC before performing subsequent processing and communication. To eliminate this conversion overhead at the power-constrained sensor node, an Analog-to-Time Converter Transmitter (ATC-Tx) directly converts analog pixel voltages into pulse-width-encoded time-domain signals. Implemented in 65-nm CMOS, the ATC-Tx supports 5~MSps with 3-bit resolution while consuming \(110~\mu\mathrm{W}\), corresponding to an energy efficiency of 21.9~pJ/sample (7.3~pJ/bit). The resulting time-domain information is transmitted using broadband Time-Domain Body Communication (TD-BC), enabling an asymmetric node--hub architecture in which energy-intensive decoding and processing are transferred to a higher-power receiver. While ATC-Tx reduces the energy required to digitize and communicate individual sensor samples, communication energy ultimately scales with the amount of information transmitted. The next part of this dissertation therefore investigates in-sensor compression before conventional digitization. A switched-capacitor-based direct analog-to-MJPEG compressor (dAJC) performs 2D-DCT, quantization, and associated JPEG processing directly on analog pixel values. The 65-nm implementation operates at 50~Mb/s while consuming 2.02~mW. By exploiting the sparsity of the compressed transform coefficients, a dynamically activated ADC digitizes only significant samples, reducing ADC conversion energy by approximately \(20\times\). This architecture demonstrates that moving compression ahead of the ADC can simultaneously reduce conversion, computation, storage, and communication overhead. Building on the direct analog compression approach, the dissertation next explores time-domain computation to improve scalability, throughput, and energy efficiency. The proposed TD-dAJC architecture represents fixed DCT and quantization coefficients using pulse-width-based time weights generated by a multi-phase PLL, while retaining the image-sensor input in the voltage domain. Multiplication is performed through time-controlled current integration, avoiding the large capacitor arrays, capacitor-ratio matching requirements, and calibration overhead associated with switched-capacitor implementations. TD-dAJC operates at up to 100~MS/s and achieves approximately 2~pJ/pixel, enabling real-time 4K video compression at 12 frames/s. The architecture provides more than an order-of-magnitude improvement in energy efficiency over the switched-capacitor implementation while also improving throughput and active-area efficiency. The dissertation then addresses the complementary challenge of supplying energy to severely constrained and battery-less biomedical systems. A dual-mode inductive--capacitive wireless power-transfer architecture is developed using temporally reconfigurable rectifiers, load-isolation techniques, and an orthogonal cuboid-based coupling structure. The architecture improves charging speed by approximately \(3.4\times\), reduces leakage during energy accumulation, and enables orientation-tolerant operation under highly attenuated implantable channels. This approach provides a compact wireless-powering solution for miniature and implantable systems where conventional batteries are undesirable or impractical. Finally, this dissertation establishes the channel conditions required for reliable multi-level time-domain communication through and around the human body. Unlike conventional communication systems in which information is primarily represented by voltage amplitude, TD-BC encodes information in signal-transition timing. Circuit models, full-wave simulations, and measurements show that channel attenuation determines edge detectability, available bandwidth determines transferable pulse width and edge formation, and group-delay variation determines temporal distortion. A 64-level pulse-width representation supports 6 bits per 100-ns symbol, corresponding to 60~Mb/s. End-to-end transmission of a 6-bit video frame through the measured body channel achieves an SSIM of 0.9791, demonstrating high-fidelity preservation of visual information despite substantial amplitude attenuation. Collectively, these works establish a system-level framework for ultra-low-power Video Sensor Nodes by reducing front-end digitization and communication energy, compressing information before digitization, exploiting time-domain representations for computation and communication, and improving wireless energy delivery. The resulting mixed-signal architectures enable scalable and energy-efficient visual sensing for next-generation wearable, biomedical, and Internet-of-Bodies systems.