Object tracking based on siamese network with multiple graph attentions

Abstract

The object tracking algorithm based on siamese network converts the object tracking problem into the problem of the similarity matching between the feature maps of the template branch and the search branch. Currently, most trackers use the entire template feature as a kernel to match the similarity of features with the search region, while this global matching approach does not cope well with rapid changes in the target. To solve this problem, we construct a graph for template and search features and embed the template feature information into the search features through multiple graph attentions. The proposed method is evaluated on advanced test benchmarks such as OTB100 and GOT10k and compared with some of the latest algorithms. Experimental results demonstrate the excellent performance of our algorithm, with significant advantages on multiple datasets.