MSCALNet: a multiscale convolutional attention LSTM network for IMU-based human activity recognition

Abstract

Wearable devices play an increasingly pivotal role in human activity recognition (HAR), particularly driven by the urgent demand in medical applications ranging from rehabilitation monitoring to fine-grained gait analysis. However, existing methods still struggle with insufficient exploration of cross-modal information, a lack of multi-timescale modeling, and the underutilization of metadata. To address these issues, we propose MSCALNet, a Multi-Scale Convolutional Attention LSTM Network. Specifically, MSCALNet employs a multi-branch differential encoding strategy to fuse heterogeneous sensor information, utilizes a joint CBAM-LSTM architecture to capture both transient and sustained activity patterns, and efficiently models multi-timescale dynamics through a dilated convolutional pyramid. Extensive experiments on three public datasets demonstrate the significant superiority of MSCALNet over state-of-the-art baselines. Furthermore, ablation studies and quantitative evaluations comprehensively validate the effectiveness of each designed module, confirming the model’s robustness and generalization capability across diverse application scenarios.