Many RGBT trackers have been proposed so far [12–15]. Due to the rapid development of RGB trackers, current RGBT trackers mainly consider the problem of dual-modal information fusion within mature trackers finetuned on the RGBT tracking task, where the key is to fuse visible and infrared image information. Several fusion methods are proposed, which are categorized as image fusion, feature fusion and decision fusion. For image fusion, the mainstream approach is to fusion image pixels based on weights [16,17], but the main information extracted from image fusion is the homogeneous information of the image pairs, and the ability to extract heterogeneous information from infrared-visible image pairs is not strong. At the same time, image fusion has certain requirements for registration between image pairs, which can lead to cumulative errors and affect tracking performance. Most trackers aggregate the representation by fusing features [18,19]. Feature fusion is a
more advanced semantic fusion compared with image fusion. There are many ways to fuse features, but the most common way is to aggregate features using weighting. Feature fusion has the potential of high flexibility and can be trained with massive unpaired data, which is well-designed to achieve significant promotion. Decision fusion models each modality independently and the scores are fused to obtain the final candidate. Compared with image fusion and feature fusion, decision fusion is the fusion method on a higher level, which uses all the information from visible and infrared images. However, it is difficult to determine the decision criteria. Luo et al. [12] utilize independent frameworks to track in RGB-T data and then the results are combined by adaptive weighting. Decision fusion avoids the heterogeneity of different modalities and is not sensitive to modality registration. Finally, these fusion methods can also be used complementarily. For example, Zhang [11] used image fusion, feature fusion and decision fusion simultaneously for information fusion and achieved good results in multiple tests. 

Transformer originates from natural language processing (NLP) for machine translation and has been introduced to vision recently with great potential [8]. Inspired by
the success in other fields, researchers have leveraged Transformer for tracking. Briefly, Transformer is an architecture for transforming one sequence into another one with the help of attention-based encoders and decoders. The attention mechanism can determine which parts of the sequence are important, breaking through the receptive field limitation of traditional CNN networks and capturing global information from the input sequence.
However, the attention mechanism requires more training data to establish global relationships. Therefore, Transformer will have a lower effect than traditional CNN networks in some tasks with smaller sample size and more emphasis on regional relationships [20]. Additionally, the attention mechanism is able to replace correlation filtering operations in the Siamese network by finding the most relevant region to the template in the search area in a global scope. The method of [9] applies Transformer to enhance and fuse features in the Siamese tracking for performance improvement.

Currently, there are few visible-light-infrared object tracking algorithms available for drones, mainly due to two reasons. Firstly, there is a lack of training data for visible -infrared images of drones. Previously, models were trained using infrared images generated from visible images due to the difficulty in obtaining infrared images. With the emergence of datasets such as LasHeR [21], it is now possible to directly use visible and infrared images for training. In addition, there are also datasets such as GTOT [22], RGBT210 [23], RGBT234 [24], etc. available for evaluating RGBT tracking algorithm performance. However, in the field of RGBT object tracking for drones, only the VTUAV [11] dataset is available. Due to the different imaging perspectives of images captured by drones compared to normal images, training algorithms with other datasets does not yield good results. Secondly, existing algorithms have slow running speeds, making them difficult to use directly. Existing mainstream RGBT object tracking algorithms are based on deep learning, which have to deal with both visible and infrared images at the same time, with a large amount of data, a complex algorithmic structure and a low processing speed, such as JMMAC (4fps) [25], FANet (2fps) [18], MANnet (2fps) [26]. In drone scenarios, there is a high demand for speed in RGBT object tracking algorithms for drones. It is necessary to simplify the algorithm structure and improve its speed.