Research On Object Tracking Methods Based On Discriminant Model

In the current wave of artificial intelligence development,video tracking technology plays an important role in various engineering fields such as intelligent monitoring,human-computer interaction,and intelligent transportation.Recently,video tracking technology has made great progress because of the deep application of machine learning.However,there are many problems and challenges in video tracking.In this thesis,two multi-scale tracking methods based on discriminative models are proposed for dealing with scale change of objective.The two discriminant tracking methods make use of a kernelized scale correlation filter as a key model.A linear scale correlation filter is converted into the scale filter by non-linear kernel tricks.The scale filter can effectively estimate 33 scale variations.Aiming at the multi-scale estimation problem of a tracking method based on dual linear structured support vector machine(DLSSVM),the thesis proposes a discriminative multi-scale tracker based on DLSSVM model and the scale correlation filter.The DLSSVM model is used as a target position estimation module of the tracker,and the scale correlation filter is used as a multi-scale estimation module of the tracker.When the scale correlation filter provides accurate scale estimation,the online classifier of the DLSSVM model can learn the appearance model of the target well and achieve a higher performance.In order to solve the problem of multi-scale estimation of fully-convolutional Siamese networks tracking(SiamFC),the thesis proposes a SiamFC model based on adaptive scale space.In this method,the scale correlation filter adaptively adjusts the multi-scale search space of the fully-convolutional Siamese networks.Then the Siamese networks discriminates the position and the scale of a target from multi-scale search space.The method indirectly solves the problem of small range multi-scale detection of SiamFC,and improves the tracking performance.A large number of experiments on a popular tracking benchmark data set have been performed to prove that the two proposed tracking methods not only significantly improve the performance of the original trackers dealing with scale changes,but also make the trackers deal with other challenges better than before.