A new robust rotation estimation method is proposed

which is based on kernelized correlation filters. A rotation pyramid is sampled around the target center and an extra 1D rotation filter is trained. Then the rotation estimation problem is tackled in a tracking-by-detection way. The rotation estimation approach is generic and can be incorporated with any tracking algorithm without a rotation estimation scheme. In the experiment

a three-layerbased tracker consisting of a 2D translation filter

a 1D scale filter and a 1D rotation filter is proposed trained and demonstrated to achieve more accurate results against state-of-the-arts on challenging dataset with rotation variation.