Research On Robust Real-time Tracking Of Textureful Object

With the fast development of computer vision, object tracking has been widely used in people’s daily life and production, such as intelligent surveillance, weather prediction, defense and military, intelligent traffic and human-computer interaction. Among all the objects, the textureful object is more probably to attract people’s attention, so our research is focus on the tracking of textureful objects. However, there are many severe factors in real world environment, such as quick movement, object deformation, occlusion, clutter background and illumination change. Researchers proposed many tracking algorithms trying to solve these problems.To realize a robust real-time tracking, three algorithms are proposed for visual tracking in this paper.Firstly, a robust real-time algorithm based on multiple instance learning is proposed, whose learning rate is dynamically updated at every frame.The strong classifier updating strategy used by multiple instance learning algorithm is relatively inefficient, bringing a poor performance in efficiency. What’s more, the learning rate in multiple instance learning algorithm is constant, meaning that the object model can’t represent the object well, which depress the robustness of multiple instance learning algorithm. Hence, this paper directly choose a set of weak classifiers, as a result, the running efficiency raised greatly. In addition, the learning rate in this paper is updated at every frame, making the updating rate of object model adapted to the change of object, which enhance the robustness.Secondly, a robust visual tracking algorithm based on online discriminant feature selection and weight multiple instance learning is proposed. When training classifiers, the multiple instance learning tracking algorithm use instance bag as training sample instead of instances, ignoring the importance of instance. DWMIL endues a weight to every instance according to its importance. What’s more, DWMIL employs two features, HOG and Haar-like, to describe object, reaching a complementation. When one of the them can’t represent object well, another play a role. As a result, we reach a more robust tracking.What’more, this paper proposed a method based on SIFT to retrack the missed object. By detecting missed object using SIFT, we can catch the object after losing object.Finally, we carried out numerous experiments to test the performance of these algorithms. The results proved that the proposed algorithms have a better performance in efficiency and robustness compared with several state-of-the-art algorithms. What’s more, we can retrack the object after losing it.