Research On Robust Target Tracking Algorithm Based On Correlation Filter

In recent years,visual object tracking technology has increasingly become a research hotspot in the field of computer vision.Object tracking algorithms based on correlation filter have received extensive attention in the field of visual object tracking,because they meet the requirements of real-time tracking while also performing excellent tracking performance.However,such algorithms still face many challenges in complex and changeable tracking scenarios.Aiming at the challenging factors in complex scenes,this thesis proposes two improved algorithms based on the correlation filter tracking framework,which effectively improves the robustness of the algorithm in complex scenes such as complex background interference,large-scale deformation,occlusion,and rotation.The research results and main contributions of this thesis are as follows:1.An adaptive context-aware correlation filter target tracking algorithm is proposed.In the traditional correlation filter object tracking algorithm,due to the limitation of the cosine window and the search area,the template learns too little background information,so in scenes with large-scale deformation and complex background interference,tracking drift is easy to occur.A framework is proposed in the context-aware algorithm to allow global contextual information to be incorporated into the correlation filter tracker,but the same suppression weight is directly used instead of calculating the interference degree of the context information to the target and it cannot adapt to give background information with different degrees of interference.To this end,this thesis has made the following improvements to the traditional correlation filter object tracking algorithm:(1)The background information around the target is learned into the filter,enhancing the classification ability of the filter template for target and context background information.(2)A formula of context information interference coefficient is proposed to quantitatively evaluate the interference degree of the context information to the target.(3)The adaptive weight coefficient vector is introduced as suppression coefficient of context background area,and based on the calculation result of the context information interference coefficient formula,the context area and the adaptive weight coefficient vector are matched,so as to achieve the result that the greater the interference degree of the context information to the target,the greater the suppression degree.Finally,based on the object tracking standard dataset to verify the effectiveness of the algorithm in this thesis,the experimental results show that the success rate and accuracy of the algorithm in this thesis are improved by 5.7%and 4.3% respectively compared with the benchmark algorithm,and it has strong robustness to difficult tracking problems such as complex background interference and large-scale deformation.2.A correlation filter object tracking algorithm fused with optical flow estimation is proposed.At present,template features such as HOG or CN extracted by mainstream correlation filter object tracking algorithms are more sensitive to rotation and deformation,and the update strategy with a fixed learning rate per frame can easily cause model degradation.To this end,this thesis combines the characteristics of local corner points that can still be detected when the moving target changes in scale,rotation,or is occluded,and the following improvements are made to the mainstream correlation filter object tracking algorithm:(1)the hog feature of the region of interest is extracted as the feature input of the correlation filter tracking model.Then,the confidence level of the tracking results of each frame is detected based on the response detection module.When the tracking confidence is low,the optical flow estimation model based on Harris corner features is turned on.Meanwhile,it is combined with the output offset of the correlation filter model to determine the optimal tracking position.On the contrary,only the correlation filter model with high efficiency is used.(2)A template adaptive learning rate adjustment scheme is proposed,which selectively adjusts the learning rate according to the tracking results of the current frame to solve the problem of template degradation problem.Finally,the performance of the algorithm is verified on the object tracking standard data set.The results show that the success rate of the proposed algorithm in the occlusion scene is increased by 6%,the success rate in the rotation scene is increased by 3.4%,and the success rate in the deformation scene is increased by 2.9%.