Show simple item record

dc.contributor.authorLuo, Xingzhi
dc.description.abstractObject tracking in streaming videos is an important research topic in computer vision. In this dissertation three topics in object tracking are covered: the background modeling based foreground object detection, the multiple object tracking models and an application of the tracking algorithm in a computer vision based lumber planning system. Background maintenance is the base of most video surveillance systems. A three level background model is proposed to tackle the problems exiting in a common surveillance system. At the pixel level, several novel background models are proposed and compared with the existing background models based on the mixture of Gaussian algorithm (parametric model) and the existing nonparametric background models. The proposed parametric background model is capable of handling the sleeping person problem and the shadow problem by using a novel weighting scheme and HSV color space. At the global level, multiple background models are maintained to represent the multiple states of the background. A fast classification algorithm based on a pyramid histogram is proposed to automatically classify each frame into the nearest background model. At the semantic level, an efficient algorithm is proposed to detect unattended objects (UAO). The proposed multiple object tracking system integrates object detection and tracking in one framework. A general solution that incorporates the most recent observation in the proposal distribution is proposed using an optical flow algorithm based on multiscale elastic matching. This solution does not rely on a measurement function as required in the Kalman particle filtering algorithm, neither does it need an auxiliary tracker as in ICONDENSATION, hence is more general and easier to be incorporated in a tracking system. A novel shape model is proposed based on a kernel-based line segment matching algorithm, which incorporates a voting scheme similar to the Radon Transform and is shown to be robust in the presence of partial occlusion. The integral image and the union set operator is combined together to corporate with the efficient computation of the region likelihood. The sampling algorithm is also improved by the incorporation of a Genetic Algorithm (GA). The integration of the detection procedure with the tracking algorithm is also proposed and tested in a face tracking scenario. As an example of the applications of multiple object tracking algorithms, an automated lumber defect detection and product planing system is presented in this dissertation, in which kalman filtering algorithm is employed to track the detected defects and construct the 3D defect models for logs.
dc.subjectObject tracking in streaming videos is an important research topic in computer vision. In\r this dissertation three topics in object tracking are covered: the background modeling based\r foreground object detection
dc.subjectthe multiple object tracking models and
dc.titleStochastic models for object tracking, background modeling and their applications
dc.description.departmentComputer Science
dc.description.majorComputer Science
dc.description.advisorSuchendra M. Bhandarkar
dc.description.committeeSuchendra M. Bhandarkar
dc.description.committeeGauri S. Datta
dc.description.committeeKhaled Rasheed
dc.description.committeeHamid R. Arabnia

Files in this item


There are no files associated with this item.

This item appears in the following Collection(s)

Show simple item record