A Robust Real-time Human Activity Recognition method Based on Attention-Augmented GRU

We proposed a robust real-time human activity recognition method based on attention-augmented Gated Recurrent Unit (GRU) using radar range profile, namely Attention-Augmented Sequential Classification (AASC). We use attention mechanism to capture the temporal relationships inherent in the range profile signatures. Therefore, our model can learn long-term temporal correlation of human activity without increasing the depth or width of recurrent neural network. The attention weights are adaptively generated using features extracted by the GRU recurrent neural network. Finally, attention-augmented features are classified by Multi-layer perceptrons. Real data of picking, boxing, rasing leg and rasing hand are collected to evaluate our model. It is shown that the proposed method outperforms the conventional GRU in recognition accuracy and robustness, demonstrating the superiority in real-time activity recognition task.