Evaluation of Transfer Learning for Human Activity Recognition Among Different Datasets
Md Shafiqul Islam, Tsuyoshi Okita, Sozo Inoue · 2019
Human activity recognition is a potential area of research. For better performance, it requires significant amount of labelled data. Collecting labeled activity data is expensive and time-consuming. To solve this problem, transfer learning has been demonstrated very effective as it gathers knowledge from labeled train data of source domain and transfers that knowledge to target domain, which has little or no labeled data. In this paper, we propose unsupervised transfer learning from source dataset to target dataset, which are completely different in terms of number of users and samples. We have used Maximum Mean Discrepancy (MMD) based transfer learning model and compared with base Convolutional Neural Network (CNN) model. We have used 4 datasets for experiment. We have trained the model on a source dataset and then transferred the model to a target dataset, which has no labels to classify activities. We have found that transfer learning model has achieved better performance compared to the base model.