Edge-Deployable Dual-Branch Network with Cross-Attention for Multi-Source Gait Recognition

Human gait serves as an important indicator for neuromuscular, musculoskeletal, and cognitive health, and its accurate recognition is critical in health monitoring, rehabilitation assessment, and identity verification. Single-sensor approaches, however, often fail to capture the complex spatiotemporal dynamics of gait. To address this limitation, we propose DBTCNet, a cross-branch attention network that fuses plantar pressure and inertial measurement unit (IMU) data for robust gait recognition. Each branch independently extracts spatial and temporal features from its modality through convolutional and temporal attention modules, while a cross-branch attention mechanism explicitly models inter-modal dependencies, enhancing complementary and discriminative feature representations. Evaluations on a self-collected dataset covering 11 gait patterns show that DBTCNet achieves 97.81% accuracy, outperforming the best single-modality approach by 0.96% in F1-score. The model was further deployed on a Raspberry Pi 4B, achieving an average inference latency of 32.14 ms with 98% accuracy, meeting real-time requirements for edge-based abnormal gait classification.