Font Size: a A A

Based On Particle Filter And A Semi-supervised Video Tracking Of Multiple Classifier

Posted on:2013-09-30Degree:MasterType:Thesis
Country:ChinaCandidate:K LiFull Text:PDF
GTID:2248330395950987Subject:Computer software and theory
Abstract/Summary:PDF Full Text Request
The rapid development of camera techniques enables more and more video based detection, tracking and analysis systems which play important roles in surveillance and many other application areas. Automatic tracking pedestrians plays a major role in a lot of real situations, such as when camera on a vehicle records the road conditions in front, it is helpful to alarm the driver to avoid accidents to detect and to track the pedestrians. Besides, it is necessary to track other objects in videos in applications such as natural human machine interface, medical image processing, autonomous robots and suspicious objects alarming in public. This thesis studies the problem of tracking objects in general videos. Solving these problems is of enormous economic and social values when multimedia software and hardware rapidly becomes more and more popular.This thesis studies two categories of video tracking algorithms:the generative method and the discriminative method. The proposed tracking systems significantly improve performance than previous related algorithms. The major work and innovations are as the following:There are two serious problems that need to be addressed. First, the movement of pedestrians and cameras in onboard videos cause non-linearity and non-Gaussianity. Second, the pedestrians, vehicles and other objects occlude each other’s a lot. This thesis proposes a pedestrian tracking algorithm that fuses multiple features in the particle filter framework. The Monte-Carlo sampling in it solves the first problem. The targets’ states are predicted by first order auto-regression dynamic model. The observation model that adaptively fuses four complementary features (HSV color histogram, LBP texture, HOG shape and motion smoothness) improves the robustness of the tracking system when there exists occlusions. The experiments carried on a large scale dataset Caltech prove that the proposed algorithm improves the recall by more than20%at the same level of precision.Tracking objects in general videos requires formulating tracking as classification which simultaneously models the appearance of the objects and the backgrounds. Based on the algorithm framework of incorporating multiple classifiers system into semi-supervised learning, this thesis propose utilizing co-training and self-training to update online the two classifiers incremental support vector machine and online random Hough forest, and fusing their results to track arbitrary objects. Semi-supervised learning employs large quantity of unlabeled sequentially arriving video data to adaptively modify the two classifiers that use complementary features. The objects are represented by superpixels which reduces noises in object models by segmentation between the foreground and the background. The mean shift clustering algorithm uses the minimal amount of labels to initialize the tracking system. The rigid and non-rigid objects tracking experiments on multiple video sequences demonstrate the proposed algorithm greatly improve the performance compared to several previous algorithms.
Keywords/Search Tags:Video Tracking, Particle Filter, Semi-supervised Learning, Multiple Classifier System, Co-training, Self-training, IncrementalSupport Vector Machine, Online Random Hough Forest, Superpixel, Meanshift
PDF Full Text Request
Related items