Transformer Tracker Based on Multi-level Residual Perception Structure

Recently, Transformer networks have been used for feature extraction and calculation of similarity in object tracking. This new structure is called one stream structure, and has achieved good results. However, the one stream structure of the Transformer tracker has too many network parameters, which limits the tracking speed of the network. For this reason, this work has designed a one stream Transformer structure that uses soft split operations to significantly reduce model parameters and computational complexity. To further improve the accuracy of tracking, this work proposes a multi-level residual perception structure to enhance the feature information of the target and reduce the background feature information, thereby enhancing the fore-ground and background discrimination ability of the model. To prove the speed and accuracy of this method, this work not only compared it with algorithms using deep neural network models, but also compared it with UAV tracking algorithms using shallow networks. Experimentally, the UAV123 dataset reached the level of SOTA; the inference speed can reach 130 FPS.