Análise de desempenho de plataformas de hardware embarcado para inferência de redes neurais convolucionais em aplicações de visão computacional

Abstract

Computer vision, coupled with the use of deep neural networks, has revolutionized areas such as automation, autonomous vehicles, and IoT devices. However, the deployment of these applications on embedded systems faces challenges due to hardware limitations, such as processing capacity and energy consumption. This work aims to evaluate and compare the performance of three representative embedded hardware platforms – Raspberry Pi 400, NVIDIA Jetson Orin Nano, and Khadas VIM 4 – in the inference of convolutional neural networks for computer vision applications, using the TensorFlow Lite framework. For this, metrics such as energy consumption,CPU/GPU usage, temperature, and inference time should be analyzed, allowing the identification of conditions that favor the use of CPU, GPU, or NPU. The results of this research are expected to provide insights for practical recommendations to help developers choose the most suitable platforms for running computer vision applications on embedded systems.