Real-time keypoint-based object tracking via online learning

Object tracking is a well-studied and challenging problem in computer vision and has many practical applications. Recently many efficient keypoint-based tracking algorithms have been proposed since the promising of effective keypoint descriptors. These methods focus on predicting the object homography transformation using a geometric estimation algorithm such as RANSAC. In addition, in these approaches the object model is often trained offline. Thus, they are not adaptive to the object appearance changes. In the paper, we propose a novel keypoint-based tracking algorithm using a online boosting framework to learn the most prominent keypoints and cluster them into patterns. The object model is represented as a combination of weighted keypoint clusters and learned through a online procedure. Our method takes advantage of binary keypoint description for clustering and thus runs at real-time. The approach focuses on predicting the target object location and is adaptive to the object appearance changes. The experimental results show that our method is robuster than other state-of-the-art keypoint-based tracking algorithms on some challenging video clips.