Uma análise para sistemas embarcados de modelos de rede neural convolucional para classificação do porte de capacete de segurança

Abstract

This work addresses the creation of machine learning models for embedded systems. In this context, we propose some trained models for classifying the correct use of safety helmets and conduct a feasibility analysis of these models in terms of their size, accuracy and losses, in order to identify the best models to be deployed on various available development boards. Based on this feasibility analysis, our experimental results show that it is possible to deploy these models on different development boards, provided that we pay attention to the constraints of each hardware, such as limited memory, processing power, low energy consumption, and others. Additionally, we created a database containing approximately four thousand images, which we used to train our neural network models.