A novel feature extraction framework based on periodic features for human activity recognition using wearable sensor data

Human Activity Recognition (HAR) is widely applied in various fields, where its accuracy and robustness are crucial. However, existing methods primarily focus on extracting spatiotemporal features, failing to fully utilize the complex periodic features present in human activities. Human behavior typically involves significant periodic patterns. Failure to effectively capture these periodic features can limit the model’s performance in handling complex behaviors, thereby affecting the accuracy and generalization ability of recognition. To address this issue, this paper proposes a network framework for periodic feature extraction—Time-FWTNet. This network integrates Fourier Transform to extract global periodic features. The FWT-Block module is designed to decompose global features and extract local periodic features, enabling the HAR method to simultaneously focus on both global and local periodic characteristics. Additionally, we propose Freq-CM to decompose high-frequency and low-frequency components for targeted feature learning, while expanding the receptive field with small convolution kernels. To verify the performance of Time-FWTNet, we conduct experiments on a constructed dataset as well as three public datasets. The experimental results show that Time-FWTNet achieves average accuracies of 98.89 %, 98.81 %, 98.76 %, and 98.72 % on the AUST-HAR, USC-HAD, UCL-HAPT, and PAMAP2 datasets, respectively.