A Resource-Aware Edge AI Architecture for Intelligent Vision-Based Embedded Systems

Abstract

Edge Artificial Intelligence (Edge AI) is becoming an important paradigm of implementing intelligent vision applications on embedded systems that have limited computational resources, memory, and power as inherent constraints. The classical cloud-based vision solutions have been afflicted with high latency, network requirement, and scalability threshold rendering them inappropriate in real time embedded application. The given paper provides a resource-aware Edge AI design to be used with intelligent vision-based embedded systems with the focus on efficient inference, low-latency, and adaptability to limited resources. The presented framework incorporates lightweight deep learning models, edge-level preprocessors, and adaptable inference pipelines to maximize the performance of using small compute and memory budgets. One use case of vision-based monitoring is tested on an embedded platform to show the effectiveness of the architecture. It has been experimentally determined that the proposed approach is capable of attaining competitive accuracy, and less inference latency, memory footprint, and power consumption than the traditional cloud-dependent or heavyweight models. The paper identifies the significance of co-designing AI inference pipelines with embedded system constraints and offers a scalable base of future intelligent edge applications.