Learning Discriminative Classifier Parameter for Visual Object Tracking by Detection

In this paper, we propose an algorithm to learn the discriminative parameter vector in successive frames of a video. The learned parameter vector is used for visual object tracking. The learning method selects some positive and negative examples from example sets. Using iterative steps, one set of positive and negative example is selected in a single iteration. The weight update strategy is then used to estimate the proportionate weight of selected example set. The examples are selected based on the criterion that when the parameter vector is updated using these examples, the classification score between parameter vector and some examples will be higher. The proportionate weights are estimated to each selected example pair based on exploring different weights; using assumption that when the parameter vector is updated, it results in higher classification score between some positive examples and the updated parameter vector. The update strategy further estimates the scalar multiplier value of selected example pairs, except the first pair, considering the target appearance change and the target drift. The tracking results of the proposed discriminative parameter learning method is comparable to the state-of-the-art video trackers.