Visual IoT: Ultra-Low-Power Processing Architectures and Implications

Visual IoT is a rapidly growing usage based on rich visual sensing, processing, and analytics. One approach for addressing visual IoT challenges is to move some computation closer to the edge device where data is captured. This article begins with a description of three key implications in ultra-low-power visual edge processing: the data footprint is constrained due to SRAM power, the available power-efficient computation is limited, and the ability to process large-scale data is challenging. To explore suitable approaches, the authors review three case studies: small-scale visual recognition for digits and characters, medium-scale visual recognition for hand gestures, and large-scale visual processing requiring video summarization. They show that co-designing algorithms and architectures for ultra-low-power processing in edge devices helps address the key challenges.