Frequency domain subpixel registration using HOG phase correlation

We present a novel frequency-domain image registration technique, which employs histograms of oriented gradients providing subpixel estimates. Our method involves image filtering using dense Histogram of oriented gradients (HOG), which provides an advanced representation of the images coping with real-world registration problems such as non-overlapping regions and small deformations. The proposed representation retains the orientation information and the corresponding weights in a multi-dimensional representation. Furthermore, due to the overlapping local contrast normalization characteristic of HOG, the proposed histogram of oriented gradients - phase correlation (HOG-PC) method improves significantly the estimated motion parameters in small size blocks. Experiments using sequences with and without ground truth including both global and local/multiple motions demonstrate that the proposed method outperforms the state-of-the-art in frequency-domain motion estimation, in the shape of phase correlation, in terms of subpixel accuracy and motion compensation prediction for a range of test material, block sizes and motion scenarios.