Studies On Real-time Visual Object Tracking In Complex Environment

Visual target object tracking has always been an important yet challenging hot topic in the field of computer vision, it involves many aspect knowledge including pattern recognition, image processing, artificial intelligence, computer application and so on, with the increase of high performance computer and high quality and cheap camera, and with the increase of demand for automatic video analysis, visual target object tracking algorithms take more and more attention, in many fields of military and civilian have extremely extensive application prospect,(such as:(Intelligent monitoring system, intelligent transportation system, precision guidance system, intelligent medical diagnosis, etc.) Over the past few decades, lots of excellent visual target tracking algorithms and effective new theory have been proposed, however, due to many difficulties account for intrinsic factors (e.g. scale and pose change, shape deformation) and extrinsic factors (e.g. partial or total occlusion, illumination change, cluttered background and motion blur), in order to design an universal visual target object tracking system, which has the attribute of real-time, robust, accurate and stable, is still facing great challenges to meet the actual demand.To solve these problems, in this paper, we detailed analysis and study traditional visual target object tracking method, and generative model and the discriminative model theory as a guide. It combines academic frontiers dynamic and practical application demand, and proposes some new idea and methods to enrich visual target object tracking theory and effectively improve the accuracy and robustness of object tracking system. In this paper, we study the object tracking system mainly in complex environment with single video device yet movement. The main contents and contributions of this dissertation are summarized as follows:(1) We propose an adaptive weighted real-time compressive tracking system in co-training framework based on discriminative model. How to improve classification performance of a classifier is the main goal of the discriminative tracking method. Building a classifier, it is critical to extract the effective features to describe the target appearance. In this paper, we introduce an effective feature selection criterion in compressive tracker to eliminate redundant information. Our method adopts Anyboost functional gradient descent method to effectively choose the most discriminative features to build classifier. We use positive and negative samples to build classifier, if these samples have the same weight and not discriminative, the classification performance of classifier will be degraded to some degree. In this paper, we integrate the sample importance into compressive tracker online learning procedure and proposed a weighted method. The self-learning often suffer from drift due to it merely make use of one kind of feature to model tracking system, when the model can not effectively describe the object appearance or the tracking system has a minor error, then the appearance model ends up getting updated with a suboptimal positive example. Over time, the errors will be accumulated and can cause drift (failure). In this paper, we adopt feature fusion method to build co-training framework. We use gray feature and LBP texture feature to build two classifiers independently to update and learn each other, then we obtain an optimal tracking result by weighting classification results.(2) We propose an effective and efficient online visual tracking algorithm with an appearance model based on partial least squares and sparse learning (PLSSL) in a particle filter framework. It is prerequisite for visual target object tracking to effectively describe the object appearance. Traditional generative tracker often only adopts foreground information to model the object appearance and neglect the background information. In this paper, a low-dimensional discriminative feature subspace can be obtained via partial least squares method from high-dimensional feature space which consists of a few collected positive and negative samples. Then, we introduce popular sparse learning into PLS methods to model the candidate target object whose appearance is represented by the PLS subspace. Compared with popular sparse representation based tracking algorithms, our algorithm can process higher resolution image observations and handle challenging image sequences with drastic appearance change and background clutters. The partial occlusion problem is one of the most general yet challenging problems which remain as arguably the most critical factor for causing tracking failures. In this paper, we proposed an effective occlusion detection mechanism based on sparse representation theory framework. We make use of trivial coefficients to construct an effective online update mechanism to real-time detect and process partial occlusion or outliers...