Research On Moving Target Tracking Method Based On Correlation Filtering With Multi-layer Convolution Features

Moving target tracking is one of the hot research topics in computer vision.It is widely used in military,medical,transportation and other fields.With the development of computer hardware and software and the continuous improvement of image processing theory,the field of target tracking has been paid more and more attention by experts and scholars in recent years.The target tracking technology has achieved considerable development.The difficulty of the moving target tracking algorithm mainly comes from the changes of the target itself and background changes.Therefore,improving the performance of tracking algorithm in complex scenes is the focus of current research.The correlation filter-based tracking algorithm has developed rapidly since it was proposed and has gradually become the mainstream algorithm in the tracking field due to advantages of high speed and good robustness.The core idea of correlation tracking is to obtain the spatial response using the correlation filter trained by the feature extracted from the target,locating the target in subsequent frames.The correlation tracking algorithm can greatly improve the tracking speed since the Fourier transform is used to convert complex convolution operations into corresponding multiplications between matrix elements.However,the accuracy of the correlation tracking algorithm in complex situations is still limited,mainly due to the limited expression of traditional hand-crafted features and the unreasonable model update method.Convolutional neural networks have recently achieved great success in the fields of face recognition and image classification,and also provide new ideas for correlation tracking algorithm.Compared with traditional hand-crafted features,convolutional features extracted from convolutional neural networks can provide more target information,especially semantic information.The tracking success rate of correlation filter-based tracking algorithms in complex situations can be improved.The work carried out in this thesis includes the following aspects:(1)The theory of target tracking algorithm is studied,especially correlation tracking algorithm.The theoretical basis of the classical correlation tracking algorithm are introduced in detail,and its advantages and disadvantages are both analyzed.(2)The correlation filtering algorithm based on traditional hand-crafted features has the advantage of fast speed,but the hand-crafted feature has limited ability to express the target,especially the semantic information,which may lead to tracking failure in complex situations.Therefore,this thesis introduces multi-layer convolution features,introduces the spatial and semantic target information at the same time,designs a correlation tracking algorithm that combines multi-layer convolutional features,and conducts qualitative and quantitative tests on the public data set.(3)Model update is an important part of the correlation filter-based tracking algorithm to adapt to the change of the target appearance.The traditional method performs linear interpolation update every frame according to the fixed learning rate,and the robustness is insufficient in the case of occlusion and fast motion.In this thesis,the average peak-to correlation energy of the target response,the mean frame difference and the target displacement are used as the judgment basis to adaptively update the filter model to improve the tracking accuracy in complex situation.The algorithm proposed in this thesis effectively improves the accuracy of correlation tracking in complex situations such as occlusion,deformation and illumination,overcomes the shortcomings of traditional hand-crafted features and improves the defects of model updating.The algorithm is tested in the public test set,and the experimental results verified the tracking performance of the algorithm in complex situations.