Hierarchical Multiscale CNN With Frequency-Aware Attention for Enhanced HAR

Wearable human activity recognition (HAR) has greatly benefited from deep convolutional neural networks. However, while increasing network width or depth improves accuracy, it also leads to higher resource consumption. This necessitates a balance between recognition accuracy and computational efficiency, especially for resource-constrained devices. To address this challenge, we propose a hybrid lightweight deep learning (DL) model that optimizes this tradeoff. Our model integrates a hierarchical multiscale convolutional neural network (HMCN) for efficient feature extraction and a frequency-based attention module (FBAM) for enhanced feature representation. To maximize effectiveness, we employ various spatial domain transformations to convert inertial sensor data into activity images, which are then processed by the HMCN-FBAM model. Evaluated on five benchmark datasets, our approach achieves the state-of-the-art performance: an ${F}1$ -score of 98.74% on UCI HAR continuous wavelet transform (CWT), 99.00% on PAMAP2 (GAF), 96.87% on opportunity (CWT), 98.53% on KU-HAR (CWT), and 96.94% on FLAAP (GAF). Moreover, compared to lightweight models and existing literature, our model not only delivers superior performance but also significantly reduces computational cost, requiring only 0.15M, 0.23M, 0.39M, 0.20M, and 0.31M parameters on the respective datasets.