FTHAR: A Dual-Modal Deep Learning Model With Feature Filtering for Human Activity Recognition on Wearable Devices

Human activity recognition (HAR) based on wearable devices enables real-time monitoring of human activities via sensor data from smartphones and smartwatches, which is crucial for applications like health management and human-computer interaction. However, existing methods focus primarily on the temporal regularity of data; they neither effectively capture spatiotemporal distribution features nor exhibit robustness to noise, and overlook the feature distribution differences between dynamic and static activities. To address these issues, we propose a dual-modal deep learning model FTHAR, with its representational capabilities complementing each other. For the frequency domain branch, we first perform FF to decouple and retain key frequencies while suppressing noise, then transform 1D sensor sequences into 2D spatial tensors by SPS. Additionally, we use DSDB to pre-classify dynamic and static activities and construct dedicated classifiers. For the time domain branch, Individual Networks extract features from each sensor independently to avoid cross-sensor interference, while a Merge Network fuses multi-sensor features to capture inter-sensor correlations. Experimental results show that FTHAR significantly improves accuracy and robustness in HAR tasks. The complete source code for the proposed strategy is available at: https://gitee.com/cailinhao/idea1/.