Local Texture Features And Its Applications In Visual Object Tracking

Visual object tracking is one of core problems within the field of computer vision.The quality of the tracking, such as adaptivity, robustness and real-time processing,directly depends on the quality of the feature extraction algorithm. Therefore, this thesisfocuses on one of most widely used feature extraction methods, local texture features,and its application in object tracking. The three main works of this paper are describedas follows:Firstly, Local Binary Pattern (LBP) and its state-of-the-art variants are presented,based on which a novel texture descriptor named Completed Local Ternary Pattern(CLTP) is proposed. To make CLTP be more insensitive to rotation, CLTP HistogramFourier (CLTP_HF) features are computed from discrete Fourier transforms of uniformCLTP histograms. Comparative experiment results on various challenging texturedatabases demonstrate that: CLTP performs favorably against several state-of-the-artvariants of LBP under image rotation and illumination changes; it is shown thatconstructing CLTP Histogram Fourier features can effectively improve textureclassification rate.Secondly, traditional Mean Shift algorithm and its state-of-the-art variants arepresented, then a more robust Mean Shift tracker based on joint color-CLTP histogramis proposed. Specifically, three main contributions are made:(1)a distinctive andeffective target object model called joint color-CLTP histogram is proposed, which canrepresent more object structure features;(2)In order to reduce the interference ofbackground in target localization, corrected background-weighted mechanism isadapted to online update target object model, which can decrease the weights of bothprominent background color and texture features similar to the target;(3) backgroundmodel is updated by adapted online background update mechanism, which caneffectively reflect the changing background in real-world environment. Comparativeexperiment results on various challenging video databases demonstrate that theproposed tracker is more robust especially in case of heavy occlusions, illuminationchanges, complex background changes than several state-of-art variants of Mean Shift tracker.Thirdly, CLTP is integrated into the framework of real-time compressive tracking,which then compared with compressive tracker based on LBP and Haar features.Numerous comparative experiment results have demonstrated that compressive trackerbased on CLTP features outperforms tracker based on LBP features at most case,performs favorably against tracker based on Haar features under conditions likechanging illuminations, occlusions.