DTTrack: Target Tracking Algorithm Combining DaSiamRPN Tracker and Transformer Tracker

At present, transformer-based target tracking algorithms mainly use transformers to fuse deep convolution features, their tracking accuracy is competitive, however compared with convolutional neural networks, their tracking speed is slow. Due to the long-distance dependence characteristics, it is difficult to obtain rich local information when extracting visual features, the tracking results may become worse, or even the tracking may fail in the later tracking procedures. The partial target tracking algorithm based on the Siamese network has great advantages in extracting local information, however its tracking accuracy cannot fully reach the transformer-based target tracking algorithm. According to the characteristics of the two trackers, combining the response scores and Hamming distance which is used to calculate the similarity, then a target tracking algorithm combining DaSiamRPN and Transformer is proposed. This structure can judge whether the tracking effect of the transformer tracker has deteriorated according to the response score and the Hamming distance between the resulting frame and the initial frame during transformer tracking, in order to replace another tracker in time. The proposed method can reduce the drift and obtain higher accuracy as well. Experiments show that our tracker achieves good results on three datasets. Our method achieved 72.0%, 69.1%, and 67.1% success rates on the GOT-10k, OTB2015, and UAV123 datasets, respectively.