Attention-based Residual BiLSTM Networks for Human Activity Recognition

Human activity recognition (HAR) commonly employs wearable sensors to identify and analyze the time series data collected by them, enabling the recognition of specific actions. However, the current fusion of convolutional and recurrent neural networks in existing approaches encounters difficulties when it comes to differentiating between similar actions. To enhance the recognition accuracy of similar actions, we suggest integrating the residual structure and layer normalization into a bidirectional long short-term memory network (BLSTM). This integration enhances the network's feature extraction capabilities, introduces an attention mechanism to optimize the final feature information, and ultimately improves the accuracy and stability of activity recognition. To validate the effectiveness of our approach, we extensively tested it on three public datasets: UCI-HAR, WISDM, and KU-HAR. The results were highly encouraging, achieving remarkable overall recognition accuracies of 98.37%, 99.01%, and 97.89% for the respective datasets. The experimental results demonstrate that this method effectively enhances the recognition accuracy of similar behaviors. A codebase implementing the described framework is available at: https://github.com/lyh0625/1DCNN-ResBLSTM-Attention.