Research On Adaptive Correlation Filter For Visual Tracking Using Deep Convolutional Feature

Visual target tracking has become a challenging research topic in the field of computer vision due to many problems such as deformation,rotation,occlusion and so on.In recent years,correlation filter based visual tracking algorithms have achieved a great breakthrough in both speed and accuracy.Based on this,this paper has carried out a research on the adaptive correlation filter for visual tracking using deep convolution features.The main work and the results obtained of this research are as follows:1.Aiming at solving the change of the target's attitude such as deformation and rotation during the tracking process,a method of using convolutional neural network to get the deep convolutional featur is proposed.Based on the generalization ability of the deep network trained on large-scale database,the high-level convolutional layer of the pre-trained convolutional neural network is used as the feature extractor to the output of the middle layers are generated in forward propagation and get resampled and dimention deduction.The simulation results show that the high-level convolution feature are highly resistant to the change of the target attitude,while the low-level convolution feature retains more details of the spatial texture.2.In order to prevent of missing of spatial details of high-level convolution feature space from lowering the accuracy of object position,a cascaded structure of the correlation filter is designed.By using the high-level features to achieve the coarse position of the target as the search area of the next-level filter,the convolutional features of different layers can be fully utilized by reducing the search area level by level.The last level of the cascade structure uses low-level convolution features to build a scale detection correlation filter and exploit the rich spatial texture information in the low-level convolution feature to achieve accurate estimation of the target boundary.Simulation results show that this structure improves the accuracy by 2.8% and the coverage success rate by 3.0%.3.Aiming at the contamination of tracker model by occlusion information during target tracking,a novel method of adaptive model updating is proposed.Feature points are extracted in the first frame.The feature points are used as the local representation of the target.By establishing the feature point library,the elements in the library are updated only in the subsequent frames by using reliable matching points.For the tracking result of a new frame,the feature point library is used to calculate the matching similarity degree as a discount factor,which increases the discount rate to reduces the update learning rate adaptively when the object undergoes through occludes,and mitigates the contamination of the model.Simulation results show that this method can improve tracking accuracy by1.9% and coverage success rate by 2.1%.4.Using publicly available OTB-2013 and OTB-2015 datasets,the algorithm proposed in the paper has been validated and tested.Simulation results show that the proposed algorithm outperforms other tracking algorithms in accuracy and coverage success rate.Developed a cross-platform program and test the program compatibility under different operating systems.The testing results show that the program can fulfill efficient tracking in different platforms and different environments.