Font Size: a A A

Appearance Modeling For Visual Object Tracking

Posted on:2015-10-05Degree:DoctorType:Dissertation
Country:ChinaCandidate:Y W WuFull Text:PDF
GTID:1228330422493439Subject:Computer application technology
Abstract/Summary:PDF Full Text Request
Visual object tracking is a hot topic in the computer vision and pattern recognitioncommunity. It has many promising applications including human computer interaction,intelligent surveillance, and medical image processing, etc. In general, a typical visualobject tracking system is composed of four modules: object initialization, appearancemodeling, motion estimation, and object localization. Among these modules, appearancemodeling is one of the most critical prerequisites for successful visual tracking. Designingan effective appearance model, however, is a challenging task due to appearance variationscaused by background clutters, object deformation, partial occlusions, and illuminationchanges, etc. To deal with these challenges, in this thesis, we focus on how to design arobust appreance model for visual tracking using different visual representations and/orstatistical modeling techniques. The main contributions of this thesis include:We present a new discriminative tracker via landmark-based label propagation (LLP)that is non-parametric and makes no specific assumption about the sample distribution.With an undirected graph representation of samples, the LLP locally approximates the softlabel of each sample by a linear combination of labels on its nearby landmarks. It is able toeffectively propagate a limited amount of initial labels to a large amount of unlabeledsamples. To this end, we introduce a local landmarks approximation method to compute thecross-similarity matrix between the whole data and landmarks. And a soft label predictionfunction incorporating the graph Laplacian regularizer is used to diffuse the known labelsto all the unlabeled vertices in the graph, which explicitly considers the local geometricalstructure of all samples. Tracking is then carried out within a Bayesian inferenceframework where the soft label prediction value is used to construct the observation model.Both qualitative and quantitative evaluations on the benchmark dataset demonstrate that theproposed algorithm outperforms the state-of-the-art methods.We present a novel online discriminative tracking framework which explicitly couplesthe objectives of example collection and classifier learning. Our method uses LaplacianRegularized Least Squares (LapRLS) to learn a robust classifier that can sufficiently exploitunlabeled data and preserve the local geometrical structure of feature space. To ensure thehigh classification confidence of the classifier, we propose an active example selectionapproach to automatically select the most informative examples for LapRLS. Part of the selected examples that are labeled using a prior classifier to enhance the adaptivity of ourtracker, which actually provides robust supervisory information to guide semi-supervisedlearning. With active example selection, we are able to avoid the ambiguity introduced byan independent example collection strategy, and to alleviate the drift problem caused bymisaligned examples. Comparison with the state-ofthe-art trackers on the comprehensivebenchmark demonstrate that our tracking algorithm is more effective and accurate.To consider the geometrical and discriminating manifold structures of the data space,we propose a novel sparse representation of symmetric positive definite (SPD) matriceswhich is performed in the data manifold adaptive kernel space. The graph Laplacian as asmooth operator of the manifold is incorporated into the kernel space to discover themanifold structure. The obtained sparse representations vary smoothly along the geodesicsof the data manifold which have more discriminating power. Tracking is then carried outwithin a Bayesian inference framework, in which the similarity of sparse representationsbetween the candidate and the template is used to construct the observation model.Numerous experiments on the benchmark set demonstrate that the proposed trackerperforms favorably against several sparsity-based algorithms.We advocate an approach to visual tracking that seeks for an appropriate metric in thefeature space of sparse codes, and propose a metric learning based structural appearancemodel for more accurate matching of different appearances. This structural representationis acquired by performing multi-scale max pooling on the weighted local sparse codes ofimage patches. An online multiple instance metric learning algorithm is proposed thatlearns a discriminative and adaptive metric, thereby better distinguishing the visual objectof interest from the background. The multiple instance setting is able to alleviate the driftproblem potentially caused by misaligned training examples. Tracking is then carried outwithin a Bayesian inference framework, in which the learned metric and the structureobject representation are used to construct the observation model. Comprehensiveexperiments on the benchmark set and other challenging image sequences demonstratequalitatively and quantitatively that the proposed algorithm outperforms the stateof-the-artmethods.
Keywords/Search Tags:Object tracking, Appearance model, Label propagation, Active learning, Semi-supervised learning, Riemannian manifold, Sparse coding, Metric learning
PDF Full Text Request
Related items