Visual object tracking based on sequential learning of SVM parameter

In this paper, a training algorithm is proposed to online (sequentially) learn the SVM based parameter. The proposed online learning method is used to construct discriminative classifier for visual object tracking application, where new examples are available in each successive frame of a video. The iterative method is based on maximizing the magnitude of sum of projection values of a positive example and a negative example which are closest to the hyperplane formed by parameter to be updated. In the proposed training framework, even if there is non-accurate labeling of the training examples (which are received online), it is possible to learn the parameter that ensures maximum value from the classification function. The learned parameter along with some examples which are nearest to hyperplane are used for the construction of object likelihood model. Using likelihood model, tracking instance most similar to the target is selected, where the target candidates are generated using particle filter framework. An object representation is also learned based on sparse DCT coefficients. This representation contains the basic structure of the target appearance. The proposed object tracking method performs better than state-of-the-art trackers in a number of challenging video sequences. When high dimension feature vectors are used, instead of simple raw pixels based feature vectors, to represent the training examples, the performance of the object tracking is better in a number of video sequences even without integrating sparse DCT coefficient based object representation as an additional model.