A Comparison of Unidirectional and Bidirectional LSTM Networks for Human Activity Recognition

Human activity recognition targets identifying different classes of human movements using data gathered from various types of sensors. Deep learning approaches, such as Recurrent Neural Networks, are gaining interest in the classification of human activities using time series data. Long-Short Term Memory is a recurrent neural network approach that is well suited for the classification of time series data where it handles the vanishing gradient and the long-term dependency problems efficiently. In this paper, we compare the human activity recognition accuracy of the unidirectional and bidirectional Long-Short Term Memory models on two different datasets that represent accelerometer data. The results show that the bidirectional approach slightly enhances the recognition quality over the unidirectional approach. However, the bidirectional approach spends more time during the training, which may hinder its applicability on large datasets.