Activity Prediction Based on Deep Learning Techniques

Abstract

Studies on real-time PM2.5 concentrations per activity in microenvironments are gaining a lot of attention due to their considerable impact on health. These studies usually assume that information about human activity patterns in certain environments is known beforehand. However, if a person’s activity pattern can be inferred reversely using environmental information, it can be easier to access the levels of PM2.5 concentration that affect human health. This study collected the actual data necessary for this purpose and designed a deep learning algorithm that can infer human activity patterns reversely using the collected dataset. The dataset was collected based on a realistic scenario, which includes activity patterns in both indoor and outdoor environments. The deep learning models used include the well-known multilayer perception (MLP) model and a long short-term memory (LSTM) model. The performance of the designed deep learning algorithm was evaluated using training and test data. Simulation results showed that the LSTM model has a higher average test accuracy of more than 15% compared to the MLP model, and overall, we were able to achieve high accuracy of over 90% on average.