2D amplitude-modulation frequency-modulation - based method for motion estimation

We present a new approach to compute the motion estimation in digital videos using 2D amplitude-modulation frequency-modulation (AM-FM) methods. The optical flow vectors are computed using an iteratively reweighted norm for total variation (IRN-TV) algorithm. We compare the proposed method using synthetic videos versus a previous three-dimensional AM-FM based method and available motion estimation methods available in free and commercial software. The proposed approach produces dense estimates that are more accurate than competitive methods.

I. Introduction

The development of perception-based motion estimation methods based on multiscale amplitude-modulation frequency-modulation (AM-FM) methods has a strong interest in digital video processing [1]. The standard use of perception-based methods for motion estimation methods rely on the use of Gabor filters, requiring a large number of frames to be computed [1]. Our motivation for the new perception-based approach comes from the need to maintain compatibility with the majority of optical flow motion methods that are based on a small number of video frames [2], [3] and the fact that there are no perception-based methods that can produce motion estimation estimates at every single pixel in biomedical applications like in speckled ultrasound videos [4], [5].