Spatial graph attention network-based object tracking with adaptive cosine window

Most popular Siamese trackers optimize the classification map from the tracking head using a fixed cosine window penalty. However, this fixed operation, which sets the weight and center of the cosine window to fixed values, can lead to tracking errors when there are similar interferences or the target is out of view. In addition, traditional graph attention networks determine attention weights only based on the cosine similarity between nodes, ignoring the relationship between the positions of nodes in the template and search region. To address these issues, this paper proposes a spatial graph attention network-based object tracking with adaptive cosine window in tracking head. The adaptive cosine window combines spatial-temporal information and adjusts the cosine window, using a positional bias Kalman filter to predict the offset of the target in the search region. The location-based attention mask module considers both the similarity between nodes and their positions in the template and search region, rather than just node similarity, which reduces the impact of similar surroundings. The attention weights between nodes are constrained using a position matrix based on Gaussian functions. Extensive experiments on four challenging public datasets (GOT-10k, UAV123, OTB-100, and LaSOT) show that our tracker outperforms other state-of-the-art trackers.