Final Year Project ELEC/COEN 490

A Tool for Global Motion Estimation and Compensation
for Video Processing

Abstract

This report gives a study of three approaches for global motion estimation
to design a tool meant to estimate and compensate global motion in a video
sequence. Global motion estimation is a process to estimate the motion of the
background in a video sequence caused mainly by the camera motion. It is
an important task in variety of video processing applications, such as coding,
segmentation, mosaicing, video surveillance. These three approaches are meant
to be applied in coding but our investigation aims to apply them for video
segmentation and surveillance.

The first method of global motion estimation is based on the minimization
of the prediction error, the sum of squared di.erence (SSD), by a gradient
descent applied over a pyramid of two input images. The second one is an
improvement of the first one and consists of minimizing the sum of absolute
di.erence (SAD) instead of the SSD in the initial matching. Beside that, it
uses prediction and earlier termination. Finally, the third approach is totally
di.erent and is based on two stages: a feature tracking followed by a estimation
of the global translation parameters using a robust M-approach (or maximum-
likelihood approach) followed by a di.erential refinement stage based on an
iterative descending Gauss-Newton minimization of the SAD, which gives all
the global motion parameters. The robust M-estimator is used to distinguish
between reliable measures and outliers.

In this report we will show the results we obtained by the three algorithms
and compare them by using three criteria: the speed, the mean absolute dif-
ference error and the robustness. As will be shown, the three algorithms have
similar performance in the case of translation, but the one using prediction and
the last one based on feature tracking are faster than the first algorithm. The
feature tracking seems to be very efficient for global motion estimation applied
for video surveillance.

Keywords: Video processing, global motion estimation, compensation, 
gradient descent, prediction, affine model, M-approach estimator, feature tracking.