Infrared tracking for accurate localization by capturing global context information

The existing model’s representation and fusion of features is inadequate in TIR scenarios. Our design of a new infrared tracking algorithm is based on this problem. Specifically, we use the transformer structure for feature extraction in our model, which captures global information and establishes long-distance dependencies between features. Additionally, convolutional projection is adopted to replace linear projection in the basic transformer module, which enriches local information and reduces computational complexity. For feature fusion, we associate the global spatiotemporal information between the template and search feature using the global information association module, which helps us get more accurate target positions. Finally, we utilize a new prediction head to further increase the stability of the prediction box and improve model performance. Our method achieved excellent performance on multiple infrared datasets.