Tbanet: dual-branch fusion of temporal convolutional network and BiLSTM–BiGRU with channel attention for sensor-based human activity recognition

Human Activity Recognition (HAR) using sensor data is a vital research domain with widespread applications in healthcare, smart surveillance, and autonomous systems. Despite progress in deep learning, many models struggle to capture both short-term patterns and long-range temporal dependencies, limiting their generalizability across datasets and real-world conditions. Rather than introducing a fundamentally new neural operator, we propose a dual-branch temporal feature fusion architecture that achieves improved recognition performance through the principled integration and temporal balancing of established components. The first branch uses a customized Temporal Convolutional Network (TCN) with causal convolutions, which efficiently captures local and mid-range temporal patterns while preserving temporal order. The second branch integrates Bidirectional Long Short-Term Memory (BiLSTM) and Bidirectional Gated Recurrent Units (BiGRU) layers and applies a channel attention mechanism to emphasize the most relevant sensor channels. This integration-oriented design allows the model to extract both hierarchical and contextual temporal features in a synergistic manner, where each component addresses a specific limitation of single-branch approaches. By concatenating outputs from both branches, the model creates a unified representation that improves classification accuracy. We evaluate our model on UniMiB SHAR, PAMAP2, and WHARF datasets, where it consistently outperforms state-of-the-art baselines across all metrics, achieving accuracies 98.91https://github.com/lahirisoham2004/TBANET.