Sensors-based Human Activity Recognition with Convolutional Neural Network and Attention Mechanism

Recently, Human Recognition Activity (HAR) has received more and more attention. At present, Recurrent neural networks (RNN), long short-term memory in particular, are main approaches in HAR. However, RNN suffers from the fact that it cannot process sequences in parallel and longer sequences cannot be remembered well. Therefore, this paper applies the attention mechanism to explore the relevant time context, and proposes a new model, named DeepConvAttn. DeepConvAttn is based on a well-known deep learning model, DeepConvLSTM. These two models are compared in experiments, and the results show DeepConvAttn is better than DeepConvLSTM on two popular HAR benchmark datasets.