Transformer Object Tracking Based on Re-Parameterization Network

In order to decrease the amount of computation and improve the tracking performance, this paper proposes a Transformer tracking algorithm based on a re-parameterization mechanism. Firstly, the backbone network was redesigned based on re-parameterization technique. Multi-branch parallel structure was adopted in the training and this structure was reconstructed into a single-branch serial structure by using the re-parameterization technique in the tracking process. Secondly, the template feature map and search feature map extracted from the backbone network were strengthened by using Transformer’s multi-head self-attention layer. The cross-attention layer was used to achieve full pixel-level information fusion between feature maps to enhance the discriminant ability for tracking targets. Finally, the bounding box regression branch was trained with the latest CIoU-Loss function. Comparison with current tracking methods shows that Average Overlaprate of the proposed method reaches 0.606 when testing on dataset GOT-10k which exceeds SiamFC++ by 0.011. Based on LaSOT, the proposed method gets Success Rate (SR) 0.554, normalization accuracy 0.659 and accuracy 0.581 which has the promotion of 0.010, 0.036 and 0.034 compared to SiamFC++ respectively.