|M.Sc Student||Gal Aharon|
|Subject||Tracking of Sequences Using Statistically Based Active|
Contours with Unscented Kalman Filtering
|Department||Department of Electrical Engineering||Supervisor||Mr. Allen Robert Tannenbaum|
|Full Thesis text - in Hebrew|
In this project, we address some key aspects of the visual tracking problem of deformable object. We employ a combination of active contours and the unscented Kalman filter.
The problem of tracking deforming moving objects is very important with applications ranging from homeland security systems to surveillance products. Many approaches have been developed during the last few years with various advantages and disadvantages for the given tracking scenario. Some of the key methodologies employ difference images, optical flow, and active contours.
In this thesis, we have chosen to divide the problem into two, more manageable sub-problems. First, we detect the objects, and their elastic deformations in time, and then we track the rigid movements of the object. Detecting the object, at any frame of the video will be using active contours (used as a powerful segmentation technique), while tracking the object's dynamic movements will be done using the unscented Kalman filter.
The active contour approach constitutes is a very powerful tool for detecting a specified target, and segment it out from the background even in a very noisy tracking environment.
We will employ a global statistically based energy functional for this purpose. This provides a robust segmentation method which works well in clutter. The unscented Kalman Filter adds the prediction step, and fits in quite naturally with our chosen active contour model. We have tested our method on several challenging scenarios, and it showed a reliable and robust behavior.
A number of techniques are available to meet the different scenarios. Among the many algorithms there are a few that are more commonly used. Motion detection tracks the area and subtracts images. Correlation tracking samples a patch of the target and attempts to match this to the most similar patch in a reasonable region surrounding the original target position. Centroid tracking assumes that the target has its own intensity; Active contour tracking segments the target from background in each frame.
The algorithms mentioned above are useful when tracking objects of limited motion. However the tracking of objects with more complex motions is a more difficult problem. In this case it is necessary to add a dynamic tracking filter that can predict the next position of the target. In most algorithms the linear Kalman filter or its modifications are used.