Stabilization of Video Sequences


Outline of Research

This research work deals with the restoration of color video sequences 
by digital image processing techniques. In particular, image degradations 
known as 'jitter' (fast and slow), introduced into video sequences by 
unintended camera motion, are considered. The processed video 
sequences become more stable and, thus, are more suitable for viewing 
and subsequent machine interpretation.

	A model of resulting geometric degradations was developed and 
simplified for processing speed-up. A method for the specification of 
image areas (targets) as reference points for the stabilization process 
was developed. An efficient search algorithms for the targets in a video 
frame was developed which uses a hierarchical search algorithm and 
exploits frame-to-frame correlation in natural video sequences by use of 
a motion predictor. The stabilization is carried out in a two-phase  
procedure. The first learning phase sets up a list of geometric restoration 
transformations obtained by searching the reference targets frame-by frame 
through the video sequence, followed by geometric triangulation procedures. 
The second restoration phase applies the restoration data to every video 
frame of the sequence. The two-phase approach allows corrective operator 
intervention in the critical first learning phase where targets can easily be 
lost by the searching algorithm due to video content or deficiencies. The 
time-consuming restoration phase can then proceed reliably with correct 
restoration data.

	The video stabilization was implemented as a computer program 
and evaluated on an image processing workstation equipped with a Targa 32+ 
video frame grabber and a Panasonic LQ4000 recordable video disk player 
located in our Artificial Intelligence Laboratory.
	Participating in the project were Dr. Bernd J. Kurz (professor), and 
Christophe Colas (formerly M.Sc.CS student).

	This research was carried out in collaboration with TMT Production 
Ltd., Fairvale, N.B. The classification tool is to be licensed to TMT 
Productions Ltd.

Project Status: Started:   December 1994
                Completed: December 1995

Relevant Publications

Christophe Colas, Video Stabilization and Tracking, M.Sc.CS thesis, Fac. 
  of Computer Science, Univ. of New Brunswick, October 1995.
  (Thesis is classified and on embargo until April 1996).


Back to: Kurz's Research Page | Kurz's Home Page

Last revised: 6 June 1997 by Bernd Kurz bjkurz@unb.ca