Visual Tracking Based On Feature Extraction

Object tracking occupies an important position in the field of computer vision, with the main purpose of tracking the change of the target’s appearance in the video. This paper focuses on introducing the visual features used to construct the target appearance model. On this basis, two tracking methods are proposed from two aspects of generate and discriminative algorithms in this paper respectively, that is, the sparse tracking method based on gradient texture feature and the discrimination tracking method based on color statistical characteristics. Finally, Matlab integrated development environment is applied to design and implement a visual tracking system based on feature extraction.This paper first presents a novel sparse tracking method based on the gradient texture feature from mean and Gaussian curvature of the image to solve the problem caused by rotation and illumination variation. Gradient texture feature possesses rotation invariance which can handle the problems caused by rotation and illumination variation. Firstly, the gradient texture features of the templates and candidates are extracted. Secondly, each candidate is represented sparsely as a linear combination of the dictionary template atoms and reconstructed by the sparse coefficient. Finally, the candidate with the minimum reconstruction error is considered as the tracking result in the particle filter framework. Furthermore, the dictionary template is updated dynamically by incremental subspace learning method, which can effectively adapt the templates to the appearance changes of the object with less drift and reduce occlusion influence.Gradient texture features represent the object mainly from details, while performing worse in dealing with the problems caused by fast motion, motion blur and background clutter. In addition, the cost of generate algorithms is higher than discriminative algorithms. To solve aforementioned problem, a simple yet robust discriminative algorithm based on color statistical characteristics is presented. Color statistical characteristics not only effectively possess a certain number of illumination invariance but also maintain high discriminative power. We establish an affine kinematics model and keep optimizing the parameters during tracking to solve scale variation and view angle change. In addition, we employ a low-dimensional variant of color statistical characteristics handled by PCA to describe appearance changes of the target. Furthermore, color statistical characteristics are applied to train naive Bayes classifiers and update classifiers online. We consider the sample with the maximum confidence as tracking result.Numerous comprehensive experimental evaluations with state-of-the-art algorithms demonstrate the effectiveness of the proposed tracking algorithm. On this basis, Matlab integrated development environment is applied to design and implement a visual tracking system based on feature extraction. Users can track video sequences in the form of software, and can simply analysis the results of tracking.