A Hybrid Deep Neural Networks for Sensor-based Human Activity Recognition

Human activity recognition is playing a significant role to health-care. Existing methods to classify human activity are mainly based on machine learning algorithm combined with features selection method and deep neural network models. Due to the complexity and diversity of sensors signal data of human activities, there exists challenge to design a suitable model. In this paper, we designed a hybrid deep neural network model, which makes use of several kinds of deep neural networks to learning features of sensors data for human activities. However, the periodicity and transferability of human action require model is able to know spatial-temporal information and the vanishing gradient problem need the rationality for a design of model. In view of the above-mentioned factors, proposed model is designed by convolutional layers, bidirectional LSTM layers and attention modules. Experimental results on UCI-HAR dataset show, proposed model achieved an accuracy of up to 95.58%, which is the best performance among the classic classification methods.