Visual Tracking Based On Correlation Filters

Object tracking is one of the most important research topics in computer vision,applied to person monitoring,medical image understanding,unmanned driving and other fields.In single object tracking,the state of unknown target is given in the first frame.The target is to be tracked in the subsequent frames.There are challenging problems in the tracking process,such as severe occlusion,pose and scale variation,and the illumination.In order to overcome these difficulties,many excellent algorithms have been proposed.In this thesis,a structured model is established,which combines local information and global information.Deep neural network is used to extract semantic features to improve the robustness of the algorithm.Correlation filters is widely used in tracking because of its accuracy and speed advantage.However,correlation filters based on target template that causes the poor occlusion handing.The samples are constructed by the circulant matrix that limits risk of training samples due to overfitting.False negative samples reduce the robustness to background clutter.Our analyze the shortcomings of qualitative and quantitative correlation filters.In convolutional neural networks,the information contained in different convolutional layers is different.The shallow networks mainly contain the spatial properties of objects,and the relatively high-level networks contain higher-order semantic information.Our analyze the average spatial energy distribution of the correlation filters in the first frame and discover that the average energy distribution of the filters in different regions is various.In this paper,the space mask is used to force the filter to focus on different regions of the target.At the same time,two kinds of space models are constructed by using the average energy distribution of different areas.The two spatial models have complementary relationships in the space,thus constructing a spatially structured model.Our do not only consider the spatial layout of the target,but also greatly reduce the parameters of the model.The final target state is determined by the fusion of the two models.The model is updated with the average peak correlation energy.Our algorithm has compared with several state-of-the-arts qualitatively and quantitatively in benchmarks selected challenging video sequences.Experiments show that the algorithm of this paper is very robust to the tracking problem of different scenes.