To address the challenges of low perception accuracy and poor trajectory tracking robustness in lower-limb rehabilitation robots, which are primarily caused by the mutual interference of multi-source sensor signals in complex environments, this paper investigates adaptive fusion methods and modeling mechanisms for multi-source information. A dynamic gating-driven bio-mechanical cross-modal gait prediction model, termed Gate-TACD, is proposed. The model features a dual-branch feature extraction architecture: the physiological branch fuses motion intent from surface electromyography (sEMG) with joint static background features provided by a goniometer to capture the human bio-postural state; meanwhile, the kinematic branch extracts temporal context features from inertial measurement units (IMUs). A dynamic gating mechanism is designed to adaptively regulate the contribution weights of biological and mechanical modalities in real-time, thereby suppressing signal crosstalk and representation distortion under adverse conditions such as sensor slip. Furthermore, a physics-informed loss function based on the minimum jerk principle is constructed to constrain the predicted trajectories within the bounds of human kinematics, ensuring the continuity and smoothness of the output gait. Experimental results on the Camargo dataset across multiple terrains demonstrate that Gate-TACD achieves superior predictive performance, with the coefficient of determination R2 reaching up to 0.99 and the root mean square error (RMSE) consistently staying between 0.3°and 0.7°under normal conditions, significantly outperforming baseline LSTM and GRU models. Notably, in challenging scenarios involving sensor slip noise, the model maintains a mean RMSE of only 1.31°, representing a 16% reduction in prediction error compared to unimodal kinematic models. This research provides a high-precision and robust cross-modal perception framework for the intelligent control and real-time tracking of lower-limb rehabilitation robots.