Research On Target Tracking Based On Siamese Network And Correlation Filter

As one of the important research directions in computer vision,visual tracking is to accurately predict the position and size of the target in the subsequent frames of the video given the initial box of the target to be tracked.The failures of target tracking are easily caused by changes in the internal or external environment of the target.In order to solve the above problems,the tracking algorithm based on Siamese network and correlation filter are studied and improved in this paper.The main work of this paper is as follows:In this paper,a tracking algorithm HCT based on off-line learning and correlation filter is proposed to track.Firstly,the weighted historical frame target image is introduced into the GOTURN network model instead of the previous frame.Then a target confidence APME is proposed,which can be applied to any correlation filtering tracking algorithm to update the model with high confidence.Finally,enhanced GOTURN and KCF are merged into a new algorithm HCT.The fusion method is to run the enhanced GOTURN algorithm and the KCF algorithm separately,then the target is detected in the enhanced GOTURN tracking result box using KCF algorithm.The last step is the confidence levels of the two prediction results with KCF algorithms are calculated with APME,and the predicted result with higher confidence is selected as the tracking results.HCT algorithm can not only make full use of the robustness of offline learning network,but also effectively reduce the drift of classifier and appearance model.So it can effectively deal with the challenges of occlusion,deformation,rotation and so on.Experimental results on OTB100 dataset show that the distance accuracy and overlap accuracy can achieve excellent performance of 0.748 and 0.620.In addition,a trip-CFNet tracking algorithm with distractor-aware training is proposed on the basis of CFNet.Trip-CFNet algorithm consists of first frame branch,template branch and detection branch.This triplet network extracts the features of the first frame target,the last frame target and the current frame search region of the input tracking video.Then,the convolution features of first frame branch and template branch are input into their respective correlation filter network layer to extract the feature map of the appearance model.The two appearance models extracted are cross-correlated with the feature map of the detection branch respectively to get two response maps.This innovative network structure can take into account both the initial state and the variety of the target,making network more robust.Finally,the APME values of the response maps are calculated,and the final response map is obtained by weighted fusion of the two response maps.Meanwhile,the model updating strategy proposed in Chapter 3 is used to update the parameters in the correlation filter layer,which reduces the drift of correlation filter model.Distractor-aware training is proposed which enables the network to distinguish the intra-class interference.The experimental results on OTB100 dataset show that the distance accuracy and overlap accuracy of tripCFNet algorithm are 0.751 and 0.620,respectively.Compared with the baseline algorithm CFNet,the distance accuracy and overlap accuracy are improved by 0.036 and 0.034,respectively.