Research On Contextual Information Based Tracking Approach

As the development of machine intelligence, vision theory which is an important way to access the outside information is brought to the forefront. Computer vision has become a hot topic in both computer engineering and science fields. Visual tracking, as an important branch of computer vision, has a wide range of practical applications, such as military visual guidance, robot vision, intelligent surveillan-ce, traffic control, medical diagnosis, meteorologic analysis and so on. Designing a robust visual tracking algorithm has pratical significance. However, it is a very challenging issue to designing a general robust tracking approach, due to the complexity of environment and the variety of the object's appearance.A detection based tracking, which is called "tracking by detection" as well, is one of the most popular tracking approaches. It firstly divides the content of an image into two categories (foreground and background) by a binary classifier, and then takes the foreground region as a new target region. When the classifier can be updated, the tracker can online learn the new features to adapt to the changes of object's appearance. But it faces an inhernt problem that it easily drifts as the error accumulation during learning. In this paper, to solve the drift problem, we have introduced contextual information to tracking and proposed a multi-classifier tracking system. The system simultaneously detects the target and contextual objects, and the target position is finally determined by all these objects. The major work in this paper includes:Firstly, a tracking algorithm, that takes the neighboring objects as the providers of the possible target positions, has been proposed. The algorithm firsly analysises the relationship of motion traces between the neighboring objects and target. An online updating trace relation model is created as well. Then, a new tracking method is derived by composing the trace relation model and detection based tracking methods. The new tracking method could simultaneously use the motion information from neighboring objects and appearance information of target. Finally, the candidate positions are evaluated and a new target position is located. Secondly, a multi-regional joint tracking algorithm has been proposed, which takes the target's local sub-regions as the evidences of deciding target position. The appearance model of target is firstly re-represented as a set of several local sub-regions. Then, an online updating gaussian mixture voting model based on the analysis of the local-whole relation model. The new tracking method comes out by composing the voting model and detection based tracking method. Finally, the offline and online classifiers are included in proposed algorithm and the adaptability and thus stability of tracking are united in one method.