Rethink Object Tracking Based on Transformer

Most visual tracking networks with transformers use attention mechanisms instead of convolutional structures. Few new methods have focused on comparing objects and search regions in visual tracking networks. However, the association information between the target in the first frame and the search area in subsequent frames will significantly affect the predicted results of subsequent targets. As a result, we propose a tracking structure: TEDTrack, which makes full use of the correlation between a given target and the search area of the following image. TED (transformer encoder-decoder, TED) is a feature fusion mechanism of TEDTrack, which mainly uses a transformer codec to integrate the extracted template features and the next frame search region features. It can achieve better results than the Previous single-target tracking algorithm which adopted the Siamese network. The tracker in this paper has conducted many experiments on multiple challenging short-term and long-term benchmark datasets. The experimental results show the effectiveness of TEDTrack.