Transcript Optical flow (motion vector) computation

Optical flow (motion vector)
computation
Course:
Computer Graphics and Image Processing
Semester:
Fall 2002
Presenter: Nilesh Ghubade ([email protected])
Advisor:
Dr Longin Jan Latecki
Dept:
Computer and Information Science,
Temple University, Philadelphia, PA-19122
Motion Analysis
1.
Three groups of motion-related problems:
Motion detection



2.
Moving object detection and location


3.
Registers any detected motion.
Single static camera.
Used for security purposes.
Determination of object trajectory.
Static camera, moving objects OR Moving camera,
static objects OR Both camera and objects moving.
Deriving 3D properties

Use of set of 2D projections acquired at different
time instants of object motion.
Object motion assumptions

Maximum velocity.

Small acceleration.
t0


t1
Cmax * dt
t2
Common motion of object points.
Mutual correspondence.
Differential motion analysis



Simple subtraction of images acquired at
different instants in time makes motion
detection possible, assuming stationary
camera position and constant illumination.
Difference image is a binary image 
subtract two consecutive images.
Cumulative difference image:
Reveals motion direction.
 Time related motion properties.
 Slow motion and small object motion.
Constructed from sequence of ‘n’ images taking first
image as the reference image.

Example
Motion in front of a security camera.
Sobel filter edge detection.
Motion Detection- Sobel filter
10 frames/second
25 frames/second
15 frames/second
15 frames/second
Optical Flow




Optical Flow reflects the image changes due to
motion during a time interval dt.
Optical flow field is the velocity field that
represents the 3D motion of object points across
a 2D image.
It should not be sensitive to illumination changes
and motion of unimportant objects (e.g.
shadows)
Exceptions:
1.
2.
Non-zero optical flow fixed sphere illuminated by a moving
source.
Zero optical flow  smooth sphere under constant
illumination, although there is rotational motion and true nonzero motion field.
Optical Flow (continued…)




Aim is to determine optical flow that
corresponds with true motion field.
Necessary pre-condition of subsequent
higher level motion processing 
stationary or moving camera.
Provides tools to determine motion
parameters, relative distances of objects in
the image etc..
Example:
t1
t2
Assumptions
Optical flow computation is based on two
assumptions:
1.
2.
The observed brightness of any object
point is constant over time.
Nearby points in the image plane move in
a similar manner (the velocity
smoothness constraint).
Optical flow computation
The optical flow field represented in the form
of Velocity vector:
Length of the vector determines the magnitude of
velocity.
 Direction of the vector determines the direction of
motion.

Global optical flow estimation—
Local constraints are propagated globally.
 But errors also propagate across the solution.

Local optical flow estimation—
Divide image into smaller regions.
 But inefficient in the areas where spatial gradients
change slowly  here use global method,
neighboring image parts contribute.

Forms of motion
Translation at constant
Set of parallel motion
distance from the observer. vectors.
Translation in depth
relative to the observer.
Set of vectors having
common focus of
expansion.
Rotation at constant
distance from view axis.
Set of concentric motion
vectors.
Rotation of planar object
perpendicular to the view
axis.
One or more sets of
vectors starting from
straight line segments.
Representation
Locate the position of a pixel (row,col) in the current image by computing
shortest Euclidean distance with respect to 5-by-5 neighborhood in the next
consecutive frame.
16
15
14
13
12
17
4
3
2
11
18
5
0
1
10
19
6
7
8
9
20
21
22
23
24
Experiments
3-by-3 neighborhood
Experiments (contd…)
5-by-5 neighborhood
Experiments (contd…)
Experiments (contd…)
Applications of optical flow

Object motion detection.

Action recognition.

Active vision or structure of motion –



Reconstruction of 3D object by computing depth information.
If you have distance (depth) maps, you can reconstruct surface of the
object.
Facial expression recognition: reference
http://athos.rutgers.edu/~decarlo/pubs/ijcv-face.pdf
Thank you 