A simple tracking algorithm using multi-scale gradient feature

Despite much success has been achieved, object tracking still remains a challenging research field in computer vision, due to many factors and difficulties such as occlusion, illumination, rotation, pose variance, and intensively motion. To handle them, many classical invariant features, object appearance models, and well-designed but complex tracking frameworks have been proposed. However, they seldom achieve effectiveness and efficiency at the same time when implemented in tracking tasks. In this paper, we propose a simple but robust tracking algorithm based on a novel feature named multi-scale gradient feature, which is subject to a structural constraint that is described by a Gaussian distribution. As the constraint is very strong, we takes a generative and static strategy to model the object appearance in video frames and do not need background models nor adaptive on-line boosting methods. It could run very fast, and perform effectively and efficiently on challenging video sequences.