Development of metabolic rate prediction model using deep learning via Kinect camera in an indoor environment

Abstract

Currently, there are only a few conventional methods to measure individual factors affecting metabolic rate (MET) based on the thermal comfort of occupants in residential buildings. In this work, a deep learning based MET prediction using a Kinect camera, which is a non-contact sensor, was developed. The root mean squared error (RMSE) and the coefficient of variation (CV) of the RMSE were used as indicators to predict MET accuracy. A total of 31 subjects participated in the experiment (16 men and 15 women). The METs of eight representative activities in the ASHRAE Standard 55 were measured (lying down, sitting, cooking, walking, eating, house cleaning, folding clothes, and handling 50 kg of books). The predicted results for all eight activities were significantly high. (RMSE: 0.26, CV: 13%). Further, METs were analyzed according to the gender and body mass index (BMI). Results of the analysis based on gender reveal that METs of men are higher than those of women. Analysis based on BMI showed that MET increased with higher BMI. However, with respect to sitting and eating food, the higher the BMI, the lower is the MET. This paper suggest that, a creative method was developed herein for predicting MET in an indoor environment with fairly high accuracy. Moreover, the difference in MET considering behavior as a factor was analyzed according to gender and BMI; these results can be used to develop guidelines for more accurate thermal comfort control.