A multi-scale feature extraction fusion model for human activity recognition

Human Activity Recognition (HAR) is an important research area in human-computer interaction and pervasive computing. In recent years, many deep learning (DL) methods have been widely used for HAR, and due to their powerful automatic feature extraction capabilities, they achieve better recognition performance than traditional methods and are applicable to more general scenarios. However, the problem is that DL methods increase the computational cost of the system and take up more system resources while achieving higher recognition accuracy, which is more challenging for its operation in small memory terminal devices such as smartphones. So, we need to reduce the model size as much as possible while taking into account the recognition accuracy. To address this problem, we propose a multi-scale feature extraction fusion model combining Convolutional Neural Network (CNN) and Gated Recurrent Unit (GRU). The model uses different convolutional kernel sizes combined with GRU to accomplish the automatic extraction of different local features and long-term dependencies of the original data to obtain a richer feature representation. In addition, the proposed model uses separable convolution instead of classical convolution to meet the requirement of reducing model parameters while improving recognition accuracy. The accuracy of the proposed model is 97.18%, 96.71%, and 96.28% on the WISDM, UCI-HAR, and PAMAP2 datasets respectively. The experimental results show that the proposed model not only obtains higher recognition accuracy but also costs lower computational resources compared with other methods.