Real-Time Human Activity Generation using Bidirectional Long Short Term Memory Networks

Human activity recognition (HAR) has extensive real-life applications. Most of the current HAR researches are based on machine learning and deep learning approaches to multi-sensor data. These approaches particularly focus on feature extraction techniques and tune hyperparameters of the network. However, these techniques exclude the size of the dataset into consideration, which is a critical performance indicator. In this paper, the open-source tri-axial accelerometer WISDM dataset has been used for simulation and Bidirectional Long Short-Term Memory (LSTM) networks have been proposed for data generation. For finding out the similarity between the generated and the original data, the losses for each of the activities have been calculated. To evaluate the effect of generated data on performance, an activity classifier has been developed which shows that the generated data has helped to increase the overall accuracy by 4% having the best recognition accuracy as 94.1%.