Research On Multi-Template Correlation Filtering Tracking Based On Gaussian Prior And Elastic-Net Constraints

After years of research and development,target tracking has become one of the most important and challenging research topics in the field of computer vision.Due to the rapid development of target tracking technology,it has been widely used in many industries such as video surveillance,human-computer interaction,robot vision and military.Therefore,various tracking algorithms have emerged in the target tracking field.Due to the complexity of the actual tracking scenes,it is still the difficulty to propose a tracking algorithm with high accuracy,robustness and speed.After the tracking algorithm based on correlation filter is proposed,it transfonns the correlation operation in the time domain into the dot product operation in the frequency domain by Fourier transform,which greatly improves the speed of the algorithm.The algorithm achieves high tracking speed meanwhile maintaining high precision,which attracts the attention of a large number of researchers.As a result,a number of improved algorithms based on correlation filter tracking frameworks have emerged.In this paper,the target tracking algorithm based on correlation filter tracking framework is studied.After in-depth analysis of the target tracking framework based on correlation filtering,the tracking algorithm based on correlation filtering is improved by learning a more discriminative filter to achieve more robust tracking,adaptive change of the target response to solve the boundary effect problem,reasonable modeling of the target response with Gaussian prior restriction to achieve more reliable tracking and so on.The experimental results show that the improved algorithm proposed in this paper has a certain improvement in the accuracy and robustness of tracking.The primary researches and innovations in this paper are summarized as follows.(1)In order to improve the robustness of the tracking algorithm based on correlation filter,we replace the L2 norm with the elastic-net constraint of the learning filter in the original correlation filter tracking frame,and learn a more discriminative filter that can adapt to more tracking scene to achieve a more robust tracking.In addition,the traditional correlation filter tracking algorithm utilizes the cyclic matrix to generate the sample matrix,and exploits the cyclic shifts instead of the actual displacement,this approach exists the boundary effect problem.The cyclic displacement is different from the real displacement in the actual scene,which may bring the problem of boundary effect.To solve the boundary effect,we adaptively change the target response and jointly solve the filter and target response in each frame.The experimental results on the OTB dataset prove that the proposed adaptive target response correlation filter tracking algorithm based on elastic network constraints has a significant improvement in tracking performance.(2)Considering that the target response is independent of the detection step of each frame in a conventional tracking framework based on correlation filter,the error will propagate to the newly calculated filter and the tracker has a risk of unrecoverable drift.Based on this,we add the Gaussian prior measure to the target response,and judge whether the tracking result is credible by calculating the confidence interval of the response peak.Based on this,the local search strategy is integrated to fine-tune the tracking result to achieve more accurate target tracking.The experimental results on the OTB data set prove that the proposed Gaussian prior correlation filter tracking algorithm has a certain improvement in accuracy and robustness.(3)Since learning a more discriminative filter and reasonable modeling of the target response in the tracking algorithm based on correlation filter can improve the tracking effect.Then if both are integrated into the framework based on correlation filtering,will it further improve the tracking effect?In addition,we extend the previous single template derivation to multi-templates,and propose a multi-template elastic network-constrained Gaussian prior correlation filter tracking algorithm.The experimental results on the OTB data set prove that the accuracy and robustness of the tracking are significantly improved compared with the previous two algorithms.