Research On On-line Contour Tracking With Level Set Method

Target tracking is one of the basic research directions in the field of computer vision,and contour tracking plays an important role in the visual analysis and understandingasanimportant part of it.By contour tracking, we can obtain attitude,behavior, movement and other information of the target, which set a foundation for further behavior recognition and understanding of the high-level.At present, it has already widely applied to the field of human-computer interface, virtual reality, medical diagnosis and safety monitoring and other active aspects.Even so, the contour tracking still has many theoretical and technical problems to be solved, such as tracking the specified target need off-line training a large number of samples, lost or disappeared after the re emergence and partial occlusion problem etc. It has crucial value and good market propect in these fields.In this thesis, we aim to achieve the goal of on-line contour tracking for long periods of time and deal with the occlusion problem in visual surveillance application.Deep discussion and analysis are carried out on the following three aspects:1)In the aspect of contour tracking in the long-term, we proposed an algorithm which combines Level Set method with on-line detection method.Detector is combined with the tracker for help.we propose a method which uses the results of detector to reinitialize the contour for the next frame. It solves the problem of tracking drift and target loss, and achieves a long time contour tracking target; Also,detector results are considered in the process of classifier on-line updating,which improves the accuracy of sampling.2)In the aspect of partial occlusion problem of contour tracking, we propose a shape prior model based on non negative matrix factorization(NMF), and a weighted distance factor is proposed to control the model's on-line incremental learning. In this paper, we use a hierarchical level set tracking framework, and we propose a method based on NNLS to determine whether the shape model is need to combin with the apparent model. With our method, occlusion problem in the tracking process is solved.3)In the aspect of on-line contour tracking, online initialization and online learning mechanism are introduced in this paper. Online initialization includes the initialization of the Level Set, the initialization of the appearance model, the initialization of thedetector and shape model.. The on-line learning includes the incremental updating of the classifier, the detector and shape model.We overcome the problem that samples can not be off-line trained and can not adapt to the changes in the online tracking.