Recognition on the Edge: Performing Inference on Embedded Devices

This study investigates object detection applied to pattern recognition on embedded platforms, using computer vision AI models constrained to a minimal footprint in order to fit in the imposed deployment requirements, while meeting predefined performance figures. The aim of this study is to assess the feasibility of performing inference on specialized low-power edge devices without needing a cloud infrastructure, thereby enabling a broad range of IoT and embedded applications that could benefit from a local AI model to automate or assist the target task. Experimental results show metrics compatible with the required performance targets, thus demonstrating that the recent advancements in AI-specialized chips and model architectures can help in bringing computation back on-device, thus removing the need for external resources even for moderately complex tasks.