<p>Tracking scenarios often present challenges such as illumination changes, occlusion, and deformation, which introduce semantic discrepancies between visual and textual data. These discrepancies make it difficult to construct unified representations, hindering effective information fusion. To address these issues, we propose SemaTrack, a semantic-driven unified vision-language tracking framework. First, we design a multimodal semantic alignment mechanism (MSAM), leveraging a cross-attention mechanism and contrastive learning to align sub-objects within modalities. MSAM establishes correspondences between specific visual objects and related text descriptions, ensuring accurate capture of semantic consistency across modalities. Second, we introduce the spatio-temporal feature interaction method (SFIM), which injects language information into visual data before feature interaction, enabling early fusion of multimodal features. This process generates a language-guided template sequence and search region. Unlike traditional methods that rely on a single template, SFIM processes template sequences, effectively capturing dynamic target changes and contextual semantics over time. This significantly enhances the model’s spatio-temporal modeling and improves token selection quality. Furthermore, we propose a token memory (TM) that iteratively updates and stores core target features and semantic context, guiding SFIM during feature interaction. This dynamic memory mechanism allows better adaptation to appearance changes and occlusion in complex scenarios. Finally, SemaTrack achieves a 65.5% success rate on the large-scale TNL2K tracking benchmark, surpassing the previous state-of-the-art result (UVLTrack) by 2.8%, demonstrating superior robustness and accuracy.</p>