To satisfy the stringent requirements of the object tracking performance in the robot's learning-from-demonstration-framework

a new tracking algorithm that can deal with fast motions

occlusions

and drifts

is proposed. First

the Median-Flow method is used to predict the position-shift of the object and the Gaussian weight of each patch. Then

the search-region is modified and the object is located by the online multi-instance learning classifier. Afterwards

the likelihood of each patch is calculated. Finally

the results are combined under the Bayes framework to get the best prediction by exhaustive search and the online classifier is updated. Experiments in several commonly used test videos show that our method outperforms the other state-of-the-art tracking methods

especially for fast motion and drifts. Furthermore

the proposed method can run in real-time.