Robust Visual Tracking via an Online Multiple Instance Learning Algorithm Based on SIFT Features

This paper presented a SIFT based multiple instance learning algorithm to deal with the problem of pose variation in the tracking process. The MIL algorithm learns weak classifiers by using instances in the positive and negative bags. Then, a strong classifier is generated by powerful weak classifiers which are selected by maximizing the inner product between the classifier and the maximum likelihood probability of instances. The method avoid computing bag probability and instance probability M times, which reduces computational time. In the traditional MIL, Haar-like features are used to represent instances, which often suffers from computational load. To deal with the problem, Harris operator is introduced to determine the outstanding SIFT features for representing an instance. Combining the Harris operator and SIFT features, the number of the extracted features are seriously deduced. Finally, the proposed algorithm is evaluated on several classical videos. The experiment results show that the method performs better than the traditional MIL algorithm and weighted MIL algorithm (WMIL).