Activity Recognition through Feature Learning and Annotations using LSTM

The tremendous growth in ubiquitous sensor networks and IoT has led to Human Activity Recognition (HAR) emerging to be an exciting challenge. Machine learning approaches proposed in this direction have proved to be competent solutions. HAR data treatment in a recurrent form, and further analysis using deep networks such as RNNs or LSTM is largely unexplored. In this work, we treat data generated from a smart home setup towards HRA, as a series of events and perform feature extraction using a the popular LSTM variant of LSTMs. We use the Aruba dataset from the CASAS project [2]. We apply LSTM to extract features and perform annotations. Additionally we apply standard classification techniques to recognize and evaluate the different activities in the newly annotated data. From experimentations using our method we can achieve annotation accuracies of up to 79.5% which is 13.6% better than other state-of-the-art solutions.