Target Tracking System Based On Corner Prediction And Siamese Network

In recent years,the object tracking algorithms have made great progress with the development of deep neural networks,and have become one of the most significant research hotspots in computer vision technology.At the same time,the object tracking algorithm based on the siamese network has been widely concerned and applied in the object tracking field due to its huge speed advantage and good tracking performance,and has become the mainstream algorithm in the object tracking field.At present,most of the current target tracking algorithms implement the target scale estimation in the form of anchor.We believe that the anchor-free method in the current target detection field can be applied to the object tracking field and achieve competitive performance compared with the current target scale estimation method.In this paper,we propose a framework for real-time object tracking which is endto-end trained anchor-free siamese network.The backbone network in traditional siamese network is relatively shallow,so that the feature information extracted by the tracker is insufficient and lower accuracy.At the same time,the relatively shallow network cannot extract the deep information of the target,so we use a relatively deep hourglass network to provide richer feature representation.At the same time,this paper proposes a feature adjustment network called Feature Merge module to fuse template features,enhance the expression capabilities of template features,and improve positioning performance..We use the anchor-free method to achieve target scale estimation instead of using region proposal network,and use corner heatmap to regress target bounding boxes.In addition,in order to cooperate with the corner heatmap prediction branch to play a better role,we proposed the vector prediction branch,which makes the algorithm more robust to potential interferences.In order to verify the effectiveness of the method proposed in this paper,we selected four mainstream data sets as the training set,and then evaluated the algorithm on VOT2019 and Tracking Net.Experiments show that the performance of object tracking can be effectively improved through corner prediction and vector regression.