Adaptive Learning Algorithm for SVM Applied to Feature Tracking

The framework of Support Vector Machines is becoming extremely popular in the field of statistical pattern classification. Kalman filters have been used for long for doing tracking. In this paper we have investigated a technique which couples Kalman filter closely with the SVM. The problem of object tracking can be seen as a pattern recognition problem in which we are looking for a pattern (object to be tracked) in an image. However, because of the dynamics, this pattern might experience some changes over time. In order to keep track of the position of the pattern and to make out the desired pattern from the background (which may contain similar patterns), we must have some strong continuous time model. In this paper we have proposed an algorithm which combines the Markov property of the Kalman filter with the strong classification capability of SVM. The whole system has been tested on real life problems and we have discovered that with this framework we can track a particular object even in a frame which contains identical objects. The results are compared to that of obtained by color blob tracking which show the strength of the approach. This will be extremely useful in the environments in which we want to track a particular person when more then one person is present.