Body Temperature Prediction with Recurrent Neural Network and its Variants

A method to predict how the human body temperature changes over time is presented in this paper. In this work, classic recurrent neural network and its two variants are used to predict body temperature, and their predictions are compared to evaluate performance. To collect the data used for training, videos which save the temperature value in the form of pixel are recorded with FLIR ONE PRO LT, a thermal sensor, and frames extracted from the video are converted into numbers with optical character recognition technology. To make our method more valuable, the temperatures at different condition, like motionless and walking, are predicted. Experiment results show that classic recurrent neural network outperforms its two variants, this may because GRU and LSTM have more parameters than classic RNN, when training data are not enough, GRU and LSTM are more likely to overfit than classic RNN.