Hybrid CNN-LSTM-GRU with Attention for Human Activity Recognition
Nisha Dangol, Sazia Mahfuz · Procedia Computer Science · 2025
Human activity recognition (HAR) using smartphone inertial sensors plays a critical role in mobile health applications through the continuous, unobtrusive monitoring of daily activities. This study introduces a hybrid deep learning architecture combining Convolutional Neural Networks (CNNs) for spatial feature extraction, Long Short-Term Memory networks (LSTMs) and Gated Recurrent Units (GRUs) for capturing complex temporal dependencies and an attention mechanism to dynamically focus on informative time steps. We evaluated this model using the standardized DAGHAR benchmark, specifically on RealWorld waist data collected via smartphones. Our experiments demonstrated a strong average cross-validation accuracy which outperforms the other models as reported in the literature, thus highlighting the model’s potential for mobile health applications.