Siamese network tracking method based on shallow-medium-deep three-level feature fusion and variable rectangular window filtering

The invention discloses a Siamese network tracking method based on shallow-medium-deep three-level feature fusion and variable rectangular window filtering. The method comprises the following steps: S1, providing a shallow-medium-deep three-level feature fusion twin network; s2, extracting corresponding feature maps from the shallow network layer, the middle network layer and the deep network layer, and performing feature level fusion to obtain a complementary feature map; s3, obtaining a pair of similarity score diagrams through correlation calculation by using the complementary feature diagrams of the two branches of the twin network, and carrying out score level fusion on the similarity score diagrams to obtain a final score diagram; s4, searching a peak point in the score graph as an initial clustering center to complete clustering, and calculating the distance between the clustering center and the farthest clustering point; and S5, controlling the change of the size of the rectangular window according to