Attention-Based Convolutional Neural Network and Bidirectional Gated Recurrent Unit for Human Activity Recognition

Using the data collected by the sensors embedded in smartphones for human activity recognition has become an important research direction. In order to overcome the problem of handcrafted feature extraction in traditional machine learning, we propose an attention-based model which combines convolutional neural network (CNN) and bidirectional gated recurrent unit (BiGRU) for human activity recognition. This model can extract local features of the original data by CNN, and construct the extracted features into a time series as the input of BiGRU, the BiGRU can capture the temporal dependencies and extract the features from the input, by introducing the attention mechanism, the influence of key input can be strengthened. The features are classified by the Softmax function. The recognition accuracy rate on the Wireless Sensor Data Mining (WISDM) dataset reaches 98.5%, which is improved compared with the state-of-the-art methods.