Object Tracking Algorithm Based On Kernelized Correlation Filters

With the continuous advancement of intelligent technology,computer vision has received extensive attention and development.Object tracking is one of the most important research topic in computer vision,it plays an important role in transportation,security,military,robotics and so on.In practical applications,however,there are many factors that affect the tracking performance,such as illumination variations,deformations,environmental noises and so on.These factors make the development of object tracking technology encounter great challenges.In the research process of object tracking technology,discriminative leaning methods have made great achievement.Especially,methods based on correlation filters have shown great potential,correlation filters become one of the most important research direction in object tracking field at present.There are many tracking methods based on correlation filters have been proposed,despite this kind of methods has made great improvement compared with traditional tracking methods,it still cannot cope with the complexity of tracking scenes.For discriminative tracking,designing effective features and processing model drift are the key problems of tracking algorithm.The feature directly determines the target appearance model,and model drift is a common phenomenon of discriminative methods.Based on Kernelized Correlation Filters(KCF),this thesis starts from two perspectives which are feature integration and update strategy,improves the model features,and alleviates the problem of model drift.Taking into account that there are certain defects in each feature,a single feature cannot respond well to a variety of scene changes.Therefore,this thesis proposes a multi-feature integration method.In order to compensate for the vulnerability of the Histograms of Oriented Gradients(Ho G)feature used by the KCF algorithm to occlusions and scale changes,a strong color feature for object appearance changes is introduced.In the selection of color features,a discriminative color descriptor that can accommodate 50 color names is used in this thesis.By combining the discriminant color descriptor with Ho G,a more expressive combination feature is obtained.This feature integration method complements the information between the color features and the Ho G features,and effectively enhances the robustness of the classifier model.After observing the performance of the KCF algorithm under different scenes,it is found that the algorithm performs poorly in low resolution scenes.Analyzing this phenomenon,and combing with the characteristics of low resolution video sequences,it can be inferred that KCF algorithm with fixed update rate cannot adapt to rapid changes of the target.This thesis starts from historical sample weights,considers historical samples as a fixed-size sample set under a certain weight threshold according to the update strategy of KCF.Based on this conclusion,this thesis proposes a method of adaptive sample set.By reducing the sample set size and removing similar samples,the weight of recent samples and the purity of the sample set are improved.In addition,a strategy for evaluating the sample contribution of each frame is adopted,and the sample weight is determined based on the Peak to Sidelobe Ratio(PSR)value of the response results.This way of improving sample sets and the distribution of sample weights effectively improve the resistance of the algorithm to changes in the object appearance model.In this thesis,the proposed algorithm is evaluated on the OTB data set and compared it with the current mainstream tracking algorithm.The experimental results show that by applying feature integration and adaptive sample set,the algorithm can adapt to different tracking scenes better,alleviate the model drift problem,effectively improve the tracking performance.