Research On Long-term Target Tracking Methods Based On Deep Residual Network And Kernel Correlation Filter

Target tracking is a research hotspot in computer vision.It has been widely used in transportation,medical,military and many other fields.However,there are still many difficulties and challenges in target tracking,such as background clutter,illumination changes,occlusion,deformation and scale changes.With the wide application of deep learning and the continuous growth of data size,target tracking combined with deep learning has gradually been proposed by researchers.This thesis establishes a target tracking framework that integrates deep learning and kernel correlation filter,and conducts discussion and research in three aspects: accuracy position,scale adaptation and long-term tracking.Specific work is as follows:(1)A method combined deep residual network with kernal correlation filter is proposed.Firstly,the structure of Res Net50 is adjusted,and the adjusted network is trained by the ILSVRC-2012 largescale data set.The pre-trained network is used for image features extraction.Secondly,feature visualization and the tracking performance with different layers of the network are compared,the third and fourth convolutional layers are choosed in this thesis.Finally,the response results of each layer are given adaptive weights according to the tracking confidence,and the accurate target position is acheived.In this research technique,deep residual network is used to extract features which improves the accuracy of positioning.(2)A target scale adaptive tracking method based on binary sort tree is proposed.After finding the position of the target,a fast scale estimation module is added.First,HOG feature is improved in the process of trilinear interpolation.Then,the improved HOG feature is used to train the scale filter,which reduces the time of feature extraction and improves the real-time performance of target tracking.Secondly,based on the structure of the binary sort tree,continuous binary search is performed in the direction of the scale change of the target which can reduce the time cost of the scale module.Finally,a scale discrimination index is designed to find the appropriate target scale.Meanwhile,in order to adapt to changes in the appearance of the target during tracking,a model update strategy is used to update the position filter and scale filter.This research technique uses search strategy based on binary sort tree to realize the scale adaptation.(3)A long-term target tracking method based on double occlusion discrimination is proposed.In the process of long-term target tracking,the probability of the target being occluded is greatly increased,which affects the accuracy of the tracking algorithm and becomes an urgent problem in long-term tracking.Firstly,making full use of the context around the target,Occlusion Discrimination Factor with Context(ODFC)is designed.PSR and ODFC are combined to double discriminate whether the tagert is occluded.When PSR and ODFC are both smaller than the set threthods,the target is severely occluced.SVM can effectively solve the two-category classification problem and AdaBoost algorithm can increase the accuracy of classification.Combining these advantages,the AdaBoost-SVM classifier is used to re-detect the target.Meanwhile,when the target is severely occluded,the update of the position filter and scale filter are suspended in order to prevent the model drift caused by the introduction of background noise.This research technique solves the severly occlusion problem in long-term tracking.In this thesis,the performance of proposed algorithms are examined on OTB2015,and compared with several mainstream algorithms for qualitative and quantitative analysis.The experimental results show that the kernal correlation filter target tracking method based on deep residual network can effectively improve the accuracy of positioning.The scale adaptive target tracking method based on binary sort tree can effectively solve the problem of scale changes in tracking.The target tracking method with double occlusion discrimination and re-detection can effectively solve the target losing problem due to severe occlusion in long-term target tracking.