Human Activity Recognition Using LSTM Network with Dropout Technique

Human Activity Recognition (HAR) is one of the fields where the research is going on in the field of deep learning and machine learning. As the everyday work that we do is ubiquitous (present everywhere), it would be helpful if a model can predict the activities performed. Many unsupervised techniques were used before but with the advent of deep learning techniques and neural networks, it has become easy for us to train and test the datasets with more accuracy. Recurrent Neural Networks (RNN) will be used as we will be working on time series data. RNNs use Backpropagation through time (BPTT) algorithm. The Long Short Term Memory (LSTM) model combined with the dropout mechanism deal with issues of vanishing gradients and overfitting of the model. LSTM consists of a cell state for preserving memory and forgetting it, whereas dropout nullifies some active states, allowing other states to participate and train the data. In this project, a better strategy has been considered that is easier to implement than the existing deep learning models that are contemplated here. This method achieved a high accuracy of 92.67% on the experimented dataset.