A 65nm 21.9pJ/Sa Pixel to PWM Conversion SoC with Time-domain Body Communication for ULP Body-Worn Video Sensor Nodes with Distributed Real-Time Inference

With the growing presence of body-worn cameras, head-mounted devices, and smart glasses in the era of Internet of Bodies (loB) and Internet of Things (loT), the demand for low-power, smaller form-factor nodes for video sensing, computing, and communication with an ability for long term usage and easy integration on the body have been steadily increasing. As shown in Figure 1, conventional architectures employ front-end ADCs to convert all the sensed analog samples to digital bits (high data volume) and then compress it using DSP blocks before communication. However, such implementations consume high conversion energy [1] and generate huge data volumes that need significant energy for communicating over wireless modalities, like Wi-Fi (which consumes 50-400mW for HD video at 30fps), Bluetooth cannot yet transfer HD video), etc. Recent non-IC based video sensing solutions [2] uses pulse-width modulated backscatter system, where the analog sampled values are converted to pulse widths before modulation, using COTS components and backscatter communication, which cannot support higher bandwidth. Although it eliminates the need for front-end ADC, the solution has bigger form factor and can function with only smaller frame sizes (112 x 112 at 13 fps), while consuming significantly higher power (2.36 mW), which makes it unsuitable for high-speed video transmission around the human body. In recent years, Human Body Communication (HBC) has emerged as a promising low-power alternative for wireless communication, providing easy integration with the body and keeping the data secured close to the body. These sensor nodes are also required to get real time inference from the data received from the sensors for proper data analysis. This inferencing workload can be transferred to an On-Body processing hub which can be of higher power and larger form factor, hence, transferring the high-power processing from the sensors to the hub, enabling distributed processing of the signal. We present the first monolithic solution ATC-VSN (Analog Voltage to Time Converter Video Sensor Node) for ultra-low power energy harvester friendly VSNs, to achieve a low-power sensing and communication solution for transmitting video around the body. It exploits the time-domain data representation to eliminate the need for front-end ADC and utilizes the low power, wideband and secure HBC communication modality to transmit the pulse-width modulated analog signal through the human body to the On-Body processing hub where the real time inferencing can be realized exhibiting > 21.9 pJ/Sa energy efficiency, which is orders of magnitude improvement over state of the art techniques [2]. For body worn cameras, a design space is explored where a direct ATC image sensor output followed by ultralow power (ULP) HBC channel using time domain samples can enable order of magnitude reduction of power consumption in body worn VSNs, making ATC-based VSNs an energy-efficient replacement for ADC less implementations.