Online training of discriminative parameter for object tracking-by-detection in a video

Abstract

In this chapter, an online training algorithm to update a discriminative parameter vector is proposed. The initial discriminative parameter is obtained by training an SVM in the first video frame only. The positive example for SVM training is the initial target object, while the negative examples are cropped at some distance away from the target object. In the successive video frames, the parameter vector is updated based on the similarity score between the parameter vector and the vector corresponding to tracked object. The similarity score is measured using a Gaussian kernel. The learned parameter is used to construct a likelihood model. Using particle filter framework, a number of target candidates are cropped. The tracked object in each successive frame is the target candidate corresponding to the highest likelihood value.