Object tracking is a challenging problem in computer vision community. It is very difficult to solve it efficiently due to the appearance or motion changes of the object, such as pose, occlusion, or illumination. Existing online tracking algorithms often update models with samples from observations in recent frames. And some successful tracking algorithms use more complex models to make the performance better. But most of them take a long time to detect the object. In this paper, we proposed an effective and efficient tracking algorithm with an appearance model based on features extracted from the multiscale image feature space with data-independent basis and a motion mode based on Gaussian perturbation. In addition, the features used in our approach are compressed in a small vector, making the classifier more efficient. The motion model based on random Gaussian distribution makes the performance more effective. The proposed algorithm runs in real-time and performs very well against some existing algorithms on challenging sequences.
展开▼
