Research On Methods Of Video Object Tracking

Visual object tracking is a key task within the field of computer vision. It has been extensively concentrated and researched, since its wide applications in many fields in our daily life, such as visual surveillance, human-computer interaction, AI navigation, medical diagnose, augmented reality and virtual reality. With the rapid development of the information technology and the correlative subjects including video image processing, computer vision, machine learning, AI and pattern recognition, many new methods of visual object tracking emerge one after another, and the tracking performance has been extended increasingly. However, although many effective tracking methods have been proposed, there are still lots of difficulties, like illumination changes in environment, appearance variations and non-linear deformations of the object, camera shake, and the noise or disturbance in the background. Thus it is a very challenging work to design an accurate and robust algorithm on visual object tracking.To deal with the existing problems of visual object tracking in practical applications, the researches in this dissertation are focused on the fundamental problems of object tracking, by using the new technology from recent interdisciplinary researches. The main works and contributions of this dissertation are summarized as follows:1. To solve the problem on tracking object with appearance variation, an incremental extremely random forest (IERF) algorithm is proposed, dealing with online learning classification with streaming data, especially with small streaming data, based on the theories of data mining and machine learning. And a semi-supervised visual object tracking algorithm using IERF is also proposed, which could carry out the online learning and dynamic description on the complex background and the appearance models of the targets, in order to achieve the location of the target (or multiple targets) accurately. Recently, object tracking has been treated as a classification problem. As the kernel of the object tracking algorithm, the proposed IERF classifier can incrementally build the classifier efficiently and fast, given limited examples even a few number. The experimental results demonstrate the accuracy and robustness of the proposed visual object tracking method using IERF, under the conditions of the changes of target appearance etc.2. Finding correspondences between two sets of features is a fundamental problem in the research of visual object tracking. Based on the theories of multilinear algebra, two efficient feature matching algorithms: higher-order matching algorithm and multiple order matching algorithm are presented. The correspondent problem between two sets of features and object tracking using feature matching are solved by the higher-order matching method. And the multiple order matching method presents stronger descriptive power of consistent constraints than the individual order based methods, by combining constraints of different orders, e.g.,unary, pairwise, third order, etc. The higher-order and the multiple order data affinities are represented by a supersymmetric affinity tensor. By taking advantage of the super-symmetrical characteristic, a higher-order and a multiple order power iteration formula are derived, with an efficient sampling approach. A novel fourth-order potential and a method fusing all potentials of different orders are also presented. The experiments show that the proposed approaches highly improve the matching performance of feature correspondences compared to the state-of-the-art algorithms, and can achieve accurate object tracking with matching features.3. It is well known that combining multiple cues can significantly improve the performance of a visual object tracker. A discriminative video object contour tracking algorithm using multi-cue under Bayesian inference is proposed. A novel contour evolution energy is proposed which integrates an incrementally learnt model with a parametric snake model. This energy function is combined with a mixed cascade particle filter tracking algorithm which fuses multiple observation models for object contour tracking. The incrementally learnt model is used to describe the observation model on object appearance. Multiple order graph matching is performed between contours in consecutive frames. Bending energy due to contour evolution is modeled using a thin plate spline (TPS). Together with the energy achieved from the contour evolution, both of the above energies are taken as observation models for contour deformation; these models are fused efficiently using a mixed cascade sampling process. The dynamic model used in our tracking method is further improved by the use of optical flow. Experiments on real videos show that our approach provides high performance on object contour tracking.4. Segmentation based object tracking can achieve a higher tracking precision compared to the location based object tracking (the location of the object is expressed as the rectangle region or ellipse region). A novel online learning based two-stage video segmentation algorithm is proposed. Through combining both global and local information of video images, the segmentation is achieved by two-stage process from coarseness to fineness. In the first stage, the videos are pre-segmented by the unsupervised image segmentation method, and the coarse foreground is extracted by the detection of the classifier. During the second stage, the final optimal pixel-wise segmentation is obtained by using spatial-temporal Conditional Random Fields. Meanwhile, a balance sampling strategy and a sample-updating approach supervised by segmentation are also proposed, to improve the accuracy and stability on initialization and updating of the classifier separately. Experiments on challenging video sequences show that the proposed method highly improves the accuracy and the stability of video segmentation, compared to state-of-the-art methods.