| Visual object tracking is one of the basic tasks of computer vision,which plays an important role in many fields.In recent years,tracking algorithms based on deep learning(DL)has achieved remarkable results.Due to the excellent representation ability of deep feature,DL based tracking algorithms is more robust than traditional tracking algorithms,but its disadvantage is high time complexity.Among them,the siamese network tracker have both good speed and accuracy,so it has become one of the most popular research branches in object tracking algorithms.However,the siamese network tracker still has many shortcomings when it faces the challenges of clutter,deformation and occlusion.In view of the above problems,this paper improves it from different aspects.Aiming at the problem that the siamese network tracker can’t distinguish well under the background of clutter.Starting with the training data,this paper proposes an effective data augmentation strategy.The generalization ability of the network is improved by adding more object categories in the training data,and the feature discrimination ability is improved by adding rich negative sample pairs to the training data.Then,starting from the training mode,this paper uses the meta learning algorithm to train the siamese network tracker,so that the tracking model can better adapt to the current tracking target through only a few gradient descent,so as to significantly enhance the tracker’s discrimination ability.Aiming at the problem of the siamese network tracker is easy to lose the target when facing the challenges of deformation and occlusion.This paper proposes an object tracking method based on policy gradient algorithm.In this method,the tracking problem is regarded as a markov decision process,and a strategy is learned to decide whether to update the appearance template of the target or to re-detect the target.Thus,the tracker can adapt to the change of the appearance of the target when the target is deformed,and recover the lost target when the target is occluded.The strategy is represented by a policy network,and the decision is made according to the reliability of the tracking results.During the decision-making process,the response map is used as a state representation to analyze the reliability of the current tracking results.During the process of re-detection,a simple and effective re-detection method is proposed to filter a large number of search areas,which greatly improves the detection efficiency.Finally,a robust tracker with leading performance is proposed by combining the above improvements.The proposed tracker is evaluated on the existing tracking benchmark.The experimental results show that the performance of the proposed tracker is improved by 3.4%-5.5%on the basis of the original tracker. |
