Title: Shadow Resistant Video Tracking
1Shadow Resistant Video Tracking
- Hao Jiang and Mark S. Drew
- School of Computing Science
- Simon Fraser University
- Vancouver, BC, Canada
2Problem Statement
Traditional Contour Tracking based on motions
3Outline
- We present an invariant image model. We study
how to project an image to an invariant space,
such that the shadow can be greatly attenuated. - We present two new external forces to the snake
model and present an chordal snake model to deal
with object tracking in cluttering environment. - The first external force is based on predictive
contour - The chordal constraint based on a new shape
descriptor - Results and conclusion
4Invariant Image
n
Planckian Lighting
ai
x
Lambertian Surface
For narrow band Sensors
The responses
5Considering 3-sensor cameras, R,
G, B
Let rlog(R/G) , blog(B/G)
We get,
b
ref-1
The slope is determined by the camera sensors
Lighting
Material
r
6Invariant Image Generation
(log(r/g), log(b/g))
Projection
(r,g,b)
b
Invariant Image Generation
Camera characteristic orientation
o
r
7Camera Calibration
Take image of one scene under different
lightings
Shift the center of the log-log ratios
corresponding to each material to the origin
Stack the log-ratio vectors of each material
into a matrix A and do SVD AUDV Camera
Orientation V(,1)
Characteristic Orientation of Canon ES60
8For Real Image
Original image
Invariant image
9Inertia Snake Tracking I
- A predictive contour constraint
Inertia Term
If we choose quadratic norm for E(.,.) the Eular
Equation,
By introducing a artificial parameter t, the
equation can be solved by PDE
10A Chordal Constraint
- Now we further introduce a second constraint to
maintain the solidness of the shape of the
contour by maintaining the value of a shape
descriptor. - The shape discriptor is defined as
- d(s, )X(s)-X(s )
- where s in is the normalized length from
one point on the contour. Apparently, d(x, ) is
periodic for both s and .
s
1
1
0
11The Shape descriptor
s
1
0
1
The similarity of contour X and Y is,
In frequency domain
12The Chordal Snake
Here we use a simple version d(s)X(s)-X(s1/2)
The variational problem is
Where Y(s) is an accessory contour, d(s) is the
calculated from last video frame. The
corresponding reaction-diffusion PDE is
13The Chordal Snake
Now we set Y0(s) X0(s1/2), D0(s) C0(s1/2). It
is not difficult to prove the following Lemma and
theorem.
Lemma If Y(s,t1) X(s1/2,t1), D(s,t1)
C(s1/2,t1) then Y(s,t) X(s1/2,t) for any tgt
t1
Theorem Given the initial conditions of Y0(s)
X0(s1/2), D0(s) C0(s1/2), we have,
Predictive Constraint
Shape descriptor constraint
14Chordal Snake Tracking II
Chordal constraint
Predictive contour
Features
Smoothed predictive contour
Previous contour
Real object boundary
Initial contour
15The System
Motion Detection
Affine Motion Estimation
Fn
Motion Map
Fn-1
Invariant Image
warping
GVF
Invariant Image
Motion Detection
Contour Prediction
Cn-1
External Force
Init Contour
Pred Contour
Chordal Model
Cn
Inertia Snake Tracking
16An Example
Initial Contour
Prediction Contour
17Experiment Result
Two successive frames
Motion map in original color space Motion map
in invariant color space
18Experiment Result
Two successive frames
Motion map in original color space Motion map
in invariant color space
19Traditional Snake Model
Frame 1 Frame 2
Frame 3
Frame 4 Frame 5
Frame 6
Frame 7 Frame 8
Frame 9
20Tracking Result
Ball Sequence Hand Sequence
Baby Sequence
21Conclusion
- We present scheme to get shadow invariant image.
- We present a much more robust snake model.
- The proposed method can work well even though
there is strong distracting shadows - Current framework can be easily extended to the
cases when the object is passing casting shadows - Future Work
- Study scheme to deal with tracking in high
dynamic range environment - Study shadow resistant method for active appear
model
22Thank You! QA