Using Synthetic Data to Improve the Accuracy of Human Activity Recognition

Approaches to Human Activity Recognition are typically data driven, classifying activities performed by humans based on low level sensor data from ambient and/or wearable sensors. Collecting this data is a challenging task, being time consuming to collect and label the data, whilst also being expensive to deploy the acquisition systems. The use of synthetic data offers a potential solution to these challenges. This research is based on generating synthetic data, with the aim of improving the performance of HAR. The data used in this study is generated based on a well-known open dataset, Mobile Health (mHealth) dataset, which was collected with wearable sensors for 12 activities. Firstly, the real data was pre- processed and classification methods such as Decision Tree, Gaussian Naïve Bayes, and Support Vector Machine were applied. Three different synthetic generation techniques Synthetic Data Vault Probabilistic Autoregressive (SDV-PAR), Time-series Generative Adversarial Network (TGAN), and Conditional Tabular Generative Adversarial Network (CTGAN) were subsequently used. In comparison to real data, which achieves an accuracy score of 0.9725, the synthetic data generated by CTGAN achieved an accuracy score of 0.8373.