A Siamese tracker based on precise fusion of prior information and multi-layer features

With the rapid advancements in deep learning, Siamese trackers have made significant strides in terms of both performance and speed. Nonetheless, when confronted with the task of effectively managing dramatic object deformations, particularly when the aspect ratio of the target is presented in an extreme proportion, most trackers are prone to losing the target, resulting in tracking failure. To tackle this challenge, we created an anchor-free conjoined network predicated on a priori information and proportion fusion (SiamPIP). We discerned that the genuine target frame, provided by the initial frame, exhibits high purity, free from any noise interference. This represents the sole prior information in the entire tracking task, enabling us to effectively utilize the initial frame information to filter the template features extracted from the backbone network. Additionally, traditional Siamese models often overlook the differences between classification and regression tasks when amalgamating deep and shallow features. Classification tasks typically require more profound semantic information, while the features of shallow networks are better suited for regression tasks, necessitating precise fusion of features. Therefore, when faced with the challenge of severely deformed objects, our model can accurately distinguish the foreground from the background while preserving the accuracy of the regression boundary box. Notably, our model also performs well in the face of challenges such as motion blur. We have achieved exemplary performance metrics of 0.707 (AUC) on OTB100, 0.340 (EAO) on VOT2019, and 0.643 (AUC) on UAV123.