Learning Spatial-temporal Consistent Correlation Filter For Visual Tracking

The target tracking has a wide coverage in real life,for example,it has an important role in autonomous vehicles and augmented reality.Visual object tracking is an important research topic in the field of computer vision.Although significant progresses have been made along decades,designing a robust tracker is still a challenging problem due to critical factors in object tracking include occlusion,background clutter,illumination variation,scale variation.In recent years,Discriminative Correlation Filter(DCF)based approaches have shown impressive performance on object tracking benchmarks.So,this thesis proposes a spatial-temporal consistent correlation filter for target tracking to improve the robustness and accuracy of tracking algorithm under complex scene.The main contents and innovations of this paper are as follows:(1)We reformulate the conventional loss function.Correlation filtering based tracking model has received lots of attention and achieved great success in real-time tracking,however,the lost function in current correlation filtering paradigm could not reliably response to the appearance changes.By exploiting the anisotropy of the filter response,a new loss function is proposed to improve the overall tracking performance.(2)Combine the correlation filtering based tracking model with deep convolutional neural network(CNN),utilizes the rich features extracted from a pre-trained CNN.Features from deep convolutional layers are discriminative while preserving spatial and structural information,shows that the last convolutional layers are more beneficial for image classification.On the other hand,shallow convolutional layers provide higher spatial resolution,which is crucial for accurate target localization,indicates that the first convolutional layers are more suitable for visual tracking.Got correlation filter responses of different layers,in this work,we adaptively weighting each layer response,and the joint filter response put more emphasis on reliable confidence map.The spatial information of the target is taken full advantage of.(3)An adaptive updating template strategy is proposed.We update classifier coefficients adaptively according to the change of the target appearance rather than a constant learning rate,and the accuracy of the tracker is improved.