Title: Einfhrung in die erweiterte Realitt Bildverarbeitung:Tracking Prof. G. Klinker, Prof. B. Brgge 7. Ju
1Einführung in die erweiterte Realität
BildverarbeitungTracking Prof. G. Klinker,
Prof. B. Brügge7. Juli 2000
2General description
- Initialization phase
- Determine/place trackable features in the scene
- Generate feature descriptions (3D, 2D, )
- Get initial measurements
- Tracking phase using suitable detectors, past
user behavior (tracks), and motion models - predict feature locations at next time interval
- redect features
- update motion models according to new measurements
3Different Tracking Modalities
- Magnetic
- Mechanical
- Inertial
- GPS
- Ultrasound
- Optical (stationary or mobile camera)
- Hybrid
4Important Criteria
- Speed
- Precision
- Robustness
- Installation cost(financial, time)
- Usability
- Generality
5Comparison of Different Tracking Modalities
S P R G W Restrictions
Optical
- -- line of sight
Magnetic
- -- - no metals
Acoustic
- - multiple echoes
Inertial
- drift
CompasTilt
- -- no metals, earth
magnetism
dGPS
-- -- visible to at
least 4 satellites
Mechanic
-- - restricted motions
G Geometric range that is covered W Wireless?
S Speed, lag P Precision, repeatability R
Robustness under rapid motions
6Optical TrackingWhy not just Calibration?
- Cons (against tracking)
- Extra system complexity
- Motion hysteresis
- Pros (for tracking)
- Reduced search space
- Smooth user motion
- Rendering into the future
7Tracking loop
- 3D Motion model
- 2D Motion model(s)
- Measurements offeature f in previous i images
m (ft-i ft-1)
Estimate
ft-i ft-1
8Tracking loop
- 3D Motion model
- 2D Motion model(s)
- Measurements offeature f in previous i images
m (ft-i ft-1)
Predict
Estimate
ft-i ft-1
ft
9Tracking loop
- 3D Motion model
- 2D Motion model(s)
- Measurements offeature f in previous i images
m (ft-i ft-1)
Predict
Estimate
ft-i ft-1 ft
ft
Search
10Easily Trackable Targets
2n IDs
cr IDs
112D Feature Tracking
12Target Detection and Identification
- Find dark blobs onbright background
- Fit quadrilateralpolygons
- Find corners
- Read ID label
132D Motion Estimation
- Compute local motionvectors of everyfeature
(images n-1,n-2) - (more images toestimate higher-order motion
models)
142D Motion Prediction
- Prediction of localfeature motion forimage n
15Target Redetection
- Prediction of localfeature motion forimage n
16Recalibration
- Here no explicit 3D motion model!
- Fast
- Reacts quickly to changing user motions (eg head
shaking) - Jittery augmentations due to noisy image
calibration - Algorithmic options
- Complete recalibration
- Compute motion parameters disjunctly (e.g. first
rotation parameters, then translation parameters) - Use previous motion parameters as initial guesses
for new estimation (hill-climbing Tsai).
17Other 2D Tracking Techniques
18Feature Correlation
- unnormalized r(i,j) S t(x,y)
s(ix,jy) - normalized
St(x,y)-Ts(ix,jy)-Sij
r(i,j)
S t(x,y) - T2 S s(ix,jy) - Sij2
Image 2
Image 1
j
i
template t, mean T
search region s mean Sij (i,j)
19Gradient-Based Image Flow
Velocity field in the image plane due to the
motion of the observer, objects or apparent
motion
I(x1,t) - I(x1,t1) image gradient
I(x,t)
u dx
t
t1
v dy ...
I(x1,t)
I(x1,t1)
x1
x2
x
dx
20Gradient-Based Image Flow
- Aperture problem
- Two variables (u,v) per pixel only one constraint
Ix,y,t
flowu,v flowdx/dt,dy/dt
21Gradient-Based Optical Flow
- Smoothness assumptionThe velocity field varies
smoothly over an image.
P D
u uaverage - Ix v vaverage - Iy P Ix
uaverage Iy vaverage It D l2 Ix2 Iy2
P D
22Gradient-Based Image Flow
- Variational calculus
- Image-flow constraint flow(x,y,t) Exu Eyv
Et 0 - Smoothness constraint
- Minimize
2
2
2
2
du du dv dv
sm(x,y,t)
dx dy
dx dy dx dv
2
2
(flow(x,y,t) l smoothness(x,y,t)) dx dy
233D Tracking Techniques
24Kalman Filtering
- Physical model of 3D user (camera) motion
25Kalman Filtering
- Physical motion model Koller circular motion
tx v cos f ty v sin f f w v 0 w 0
26Kalman Filtering
- Linear filter
- Least squares weighted fit to estimate best
model parameters from sensor measurements - Recursive (linear regression)
27Kalman Filtering
y(n)
C(n)
S
z-1t
v1(n)
x(n)
S
v2(n)
F(n1,n)
v1(n) process noise x(n) state
vector F(n1,n) state transition matrix
v2(n) measurement noise C(n) measurement
matrix y(n) observed data
28Epipolar Geometry
29Echtzeit Stereosystem
30Verfolgung bewegter Objekte
Reale Welt
Sensordaten
Aktionen
Computer
Kommandos
Benutzer
Augmentations
31Bewegte Objekte mit Marken
32Model-basierte dynamische Szenenanalyse
33(No Transcript)
34n
-
k1
a bc d ef g hij k l m n o p qr s t u v w x y
z ABCDEFGHIJK LM NO PQR S TU V W XYZ