DEEP LEARNING APPROACH FOR HUMAN ACTIVITY RECOGNITION FOR MOBILE OR WEARABLE DEVICES

Abstract

In the modern age of technology when wearable devices are so popular and are used for sports, well being and healthcare, the amount of information processed is tremendous and can be used and analysed for biomedical applications. The human activities are a sequence of encoded samples in time i.e. real time series data, traditional machine learning techniques cannot exploit the temporal correlation between the input samples, resulting in unable to capture dependencies between the consecutive input data samples. Therefore, the traditional methods are unable to adapt to time series data and deep learning approach ts the picture. Convolutional neural networks (CNNs) is useful in capturing the dependencies in the input data samples but shallow features are neglected. We apply recurrent neural networks (RNNs) for developing a model for human activity recognition, as it can capture dependencies in input sequences. Long short-term memory (LSTM) recurrent neural network is the architecture for our model, it evaluates the performance on the three benchmark datasets. The design of the model overcomes the limitations present in a deep learning framework and traditional machine learning techniques. The experimental results demonstrate that our model outperforms all the other techniques