Learning Based Structured Image Model Extraction And Tracking

With the rapid development of information technology, most researchers are fo-cused on computer vision which is one of the most popular fields. Object tracking isa major problem in computer vision and also a hard problem to solve. Lots of effortshave been made and plenty of algorithms have been developed according to differentkinds of research background. However, there is no general solution and universaltheory till now.To describe the structure feature is an important problem in image analysis andobject tracking. A lot of related researches only take the local feature of the targetobjects, but not make use of the structure feature (the corresponding relationship ofdifferent objects) in the images. It is the problem this paper focused on.In this paper, structure feature is introduced into the traditional graph model, anda novel structured graph model is defined in the paper. In order to obtain a betterperformance, the proposed approach can make use of the corresponding relationshipof multiple objects, which is the edge feature of the structured graph model. The maincontributions of this paper are in two aspects.First, in the multiple object tracking problem, we define the structure feature ofthe graph model. Therefore, a structured graph matching model is established, and theproblem is regarded as structured node and edge matching between graphs generatedfrom successive frames. In essence, it is formulated as the maximum weighted bipar-tite matching problem to be solved using the dynamic Hungarian algorithm, which isapplicable of optimally solving the assignment problem in situations with changingedge costs or weights. In the proposed graph matching model, the parameters of thestructured graph matching model are determined in a stochastic learning process. Inorder to improve the tracking performance, the bilateral tracking is also used. Secondly, in tubular structure tracking problem, original dataset is transformedto a virtual superellipsoid with non-uniform density where the components of datasetand the local topology of tubular objects are equivalent to the inferred weights of thekernel-based model. Therefore, a structured 3D tubular tracking model is established,and the problem is regarded as maximizing the prediction energy. In order to improvethe performance, Gaussian process (GP) and unscented Kalman Filter (UKF) are in-troduced into the nonlinear prediction phase. GP can provide uncertainty estimatesand learn noise and smoothness parameters from dataset, and UKF can be utilized inthe tubular structure extraction problem.Extensive experimental results demonstrate that the proposed approaches not on-ly take the target feature into account, but also make use of the structure feature inthe images. Different from the traditional approaches, the structured graph model is abetter description of the whole graph. Therefore, the proposed approach gains a betterperformance.