Researches On Object Tracking Based On Kernelized Correlation Filters

Object tracking is an important research direction in computer vision area.It has a wide range of applications in both military fields such as precision guidance of weapons,and civilian fields such as intelligent video surveillance and human computer interaction.Object tracking is the process of achieving continuous state estimation of target as the scene changes continuously in a video sequence.Due to the complexity of tracking scenes,achieving a persistent and stable object tracking still faces a series of difficulties and challenges.In recent years,discriminant-based methods have received extensive attention due to their excellent performance.It regards object tracking as the classification problem of the object and the background,and finds object by judging the candidate sample.Tracking algorithm based on kernelized correlation filters attracts extensive attention,and this paper is to improve the method around the window function,multi-feature fusion and scale estimation.In order to improve the reliability of tracking,reduce the impact of irrelevant information,and prioritize the use of information that is highly discriminative,we proposes a method with “local saliency analysis”.Firstly,local saliency analysis is performed over the target with bottom-up visual saliency detection,and the saliency window is obtained.Then,a Gaussian saliency window is obtained by mixing the saliency window and the Gaussian window.Afterwards,the Gaussian saliency window is integrated into the kernelized correlation filtering algorithm.Experiments show that this method can make the model focused on the saliency area of the target,and suppress the influence of other non-saliency areas effectively,it achieves more stable and reliable tracking,and improve the success rate and accuracy of the algorithm.With the problem that a single feature can not effectively deal with multiple complex scenes and object's scale is difficult to be estimated accurately and quickly,a kernelized correlation filtering algorithm that combines multiple features and fast-scale estimation is proposed.It combines HOG and CN,and convert the scale problem in the Cartesian coordinate to the translation problem in the log-polar coordinate.Experiments show that this method can make full use of the advantages of different features to deal with complex scenes and estimate the target's scale quickly and accurately.