Uma abordagem para a estimativa de medidas antropométricas a partir de imagens 2D utilizando Machine Learning

Abstract

Anthropometric measurements are used in several areas of study and can be applied in public health as an indicator of cardiovascular risk; it can also be used as a parameter for making tailored clothes or even for body reconstruction in a nutritional follow-up. Thus, the automatic estimation of these measurements can provide an improvement in anthropometric processes, even more so if it is based on 2D images, as it represents a gain due to the low cost of implementing this technology. This work presents an approach to estimate these anthropometric measurements through images using Machine Learning and CNNs. Furthermore, during this research, an anthropometric measurements dataset was proposed containing 913 samples (500 women and 413 men) with images of human body and measurements (biceps, calf, forearm, hip, thigh and waist) and the patients’ additional information, like age, height and weight. Finally, experiments were performed to analyze the influence of patient information on the estimation of measures, as well as a gender analysis to obtain insights into the dataset. In the proposed approach, experiments using female data in predicting the forearm anthropometric measurement yielded a (MAE) of 0.778±0.083cm and (RMSE) of 1.096±0.292cm using the Elastic-Net regressor. We achieved an approximate 2% error with (MAPE) using SVR regressor with male gender data for predicting hip measurement. Lastly, this work presents satisfactory results since it has a small margin of error and superiority compared to the results of a similar approach in the literature.