Subject-independent gait activity recognition using DSAF: dual-stream IMU-EMG attention fusion with asymmetric temporal encoding and physiological complementarity weighting.

Introduction: Accurate gait recognition using wearable sensors is of significant clinical value for adaptive prosthetic control, lower-limb exoskeleton assistance, and objective rehabilitation assessment. However, subject-independent recognition remains a major challenge, as unseen individuals can exhibit highly variable limb kinematics and muscle activation patterns, and existing approaches often rely on a single sensor modality or naive fusion strategies that fail to leverage the complementary information between inertial and electromyographic signals. Methods: To address these gaps, this study proposes DSAF, a dual-stream attention fusion network that separately encodes kinematic (IMU) and neuromuscular (EMG) information, and adaptively integrates them through a physiological complementarity weighting mechanism designed for window-level modality adaptation. The framework is evaluated on the public HuGaDB dataset for eight common locomotion activities (e.g., walking, running, stair negotiation, and sit-to-stand transitions), using a leave-one-subject-out protocol to rigorously assess generalization to new users. Results: DSAF achieves 96.41% accuracy, 96.08% macro-precision, 95.62% macro-recall, and 95.81% macro-F1, consistently outperforming recent sequence-learning baselines across all 18 held-out subjects. Ablation studies further confirm that both the modality-specific encoding and the adaptive fusion mechanism contribute positively to the performance. Discussion: These findings indicate that adaptive IMU-EMG fusion can effectively strengthen wearable gait recognition, providing a promising solution for real-world rehabilitation monitoring and assistive human-machine interfaces.