Research On Moving Object Tracking Algorithm Based On Correlation Filter

The research on moving object tracking technology is an important direction in the field of computer vision,and has important applications in many fields in today's society.The moving object tracking algorithm based on correlation filter has become a research hotspot in recent years due to its advantages in tracking performance and computing cost,and has also achieved remarkable research results.However,there are many complex situations exist in the real tracking scenes,such as object deformation,background clutters,fast motion,severe occlusion and so on,which pose great challenges to the accuracy and robustness of the tracking algorithm.The thesis researches and improves the shortcomings of the current moving object tracking algorithm based on correlation filter to gain the tracking performance improvements in complex tracking scenes,the main research work of the thesis is as follows:1.Aiming at the shortcomings of current moving object tracking algorithm based on correlation filter in the training of correlation filter,multi-feature response fusion method and model update strategy,a multi-feature correlation filter tracking algorithm based on context-aware is proposed.By introducing context-aware framework,images are collected in the context region of the tracking object as negative samples for the training of the correlation filter,which can improve the robustness of the algorithm.At the same time,since the hog feature and the color histogram feature are complementary,they are selected to express the appearance of the tracking object.And the responses of the two features are fused by adjusting the response weighted factor adaptively,which improves the reliability of the fusion response map of the two features.Moreover,for the model pollution risk existing in the traditional model update strategy,the maximum response peak and the average peak-to-correlation energy are introduced as the confidence index,and the tracking model will be updated only when the high confidence update conditions are satisfied.The results of the experiments show that the proposed algorithm has better tracking accuracy and robustness in different tracking scenes.2.Considering the limitation of the traditional hand-craft image features in the expression of object appearance,the thesis uses different layers of depth features to express the object appearance,and proposes a correlation filter tracking algorithm using multi-layer depth features.The thesis chooses to extract deep feature and shallow feature from the VGGNet-19 network and train the correlation filters respectively,which can realize precise localizing of the tracking object while enhancing the discrimination ability of the algorithm.At the same time,aiming at the problem of high computational cost in the scale estimation method,the dimension of feature is reduced by using principal component analysis,and the computational cost of the scale estimation method is reduced.An object re-detection module is introduced to the proposed algorithm.When the value of the smooth constraint of confidence maps is greater than the threshold,the object re-detection is performed to restore the tracking of the object,which improves the tracking performance of the algorithm in the case of occlusion and out-of-view.The results of the experiments show that the proposed algorithm has good tracking performance in complex tracking scenes,and can maintain stable tracking of the object when dealing with the significant appearance changes of the object,severe occlusion,out-of-view and other scenes.