ABSTRACT
The recent advancements towards Artificial Intelligence (AI) at the edge resonate with an impression of a dichotomy between resource intensive, highly abstracted Machine Learning (ML) research and strongly optimized, low-level embedded design. Overcoming such opposing mindsets is imperative for enabling desirable future scenarios such as autonomous driving and smart cities. edge AI must incorporate both straightforward streamlined deployments together with resource efficient execution to achieve general acceptance. This research aims to exemplify how such an endeavour could be realized, utilizing a novel low power AI accelerator together with a state-of-the-art object detection algorithm. Different considerations regarding model structure and efficient hardware acceleration are presented for deploying Deep Learning (DL) applications in resource restricted environments while maintaining the comfort of operating at a high degree of abstraction. The goal is to demonstrate what is possible in the field of edge AI once software and hardware are optimally matched.
