TeamStation AI System Report LATAM IT Salaries 2024
CRF-Filters: Discriminative Particle Filters for Sequential State Estimation
1. CRF-F: D P F
S S E
B L, D F L L
Hannes Schulz
University of Freiburg, ACS
Feb 2008
2. O
1 I: S E U D M
2 T D M CRF
Short Introduction to CRF
CRF-Model for State Estimation
3 A
CRF-Filter Algorithm
Learning the Parameters
4 E R
3. O
1 I: S E U D M
2 T D M CRF
Short Introduction to CRF
CRF-Model for State Estimation
3 A
CRF-Filter Algorithm
Learning the Parameters
4 E R
4. Intro Transformation of Directed Model to CRF Application Experimental Results
C: S E
C D M A S E
ut−2 ut−1
xt−2 xt−1 xt
... ... n
1 2 n
zt−1 1
zt 2
zt zt
zt−1 zt−1
P (xt |u1:t −1 , z1:t ) = ηP (zt |xt ) P (xt |ut −1 , xt −1 )P (xt −1 |u1:t −2 , z1:t −1 ) dxt −1
5. Intro Transformation of Directed Model to CRF Application Experimental Results
D D M P
p (zt |xt ) = n
i =1 p (zti |xt ) p (xt +1 |xt , u)
6. Intro Transformation of Directed Model to CRF Application Experimental Results
D D M P
p (zt |xt ) = n
i =1 p (zti |xt ) p (xt +1 |xt , u)
i
P (zt |xt) ˆi
zt zmax
zrand
7. Intro Transformation of Directed Model to CRF Application Experimental Results
D D M P
p (zt |xt ) = n
i =1 p (zti |xt ) p (xt +1 |xt , u)
i
P (zt |xt) ˆi
zt zmax
zrand
8. Intro Transformation of Directed Model to CRF Application Experimental Results
D D M P
p (zt |xt ) = n
i =1 p (zti |xt ) p (xt +1 |xt , u)
i
P (zt |xt) ˆi
zt zmax
δrot2
zrand
xt
δtrans
δrot1
xt−1
u = (δrot1 , δrot2 , δtrans )
executed with gaussian
noise
9. Intro Transformation of Directed Model to CRF Application Experimental Results
A P D A
p (zti |xt ) are not cond. independent
zt
xt
10. Intro Transformation of Directed Model to CRF Application Experimental Results
A P D A
ut−2 ut−1
xt−2 xt−1 xt p (zti |xt ) are not cond. independent
Sensor models can only be
... ... n
1 2 n
zt−1 1 2 zt
zt−1 zt−1 zt zt generated seperatly for each beam
i
P (zt |xt) ˆi
zt zmax
zrand
11. Intro Transformation of Directed Model to CRF Application Experimental Results
A P D A
ut−2 ut−1
xt−2 xt−1 xt p (zti |xt ) are not cond. independent
Sensor models can only be
... ... n
1 2 n
zt−1 1 2 zt
zt−1 zt−1 zt zt generated seperatly for each beam
Assumption that measurements
are independent: “Works
i
P (zt |xt) ˆi
zt zmax
surprisingly well”. . . if. . .
zrand
12. Intro Transformation of Directed Model to CRF Application Experimental Results
A P D A
ut−2 ut−1
xt−2 xt−1 xt p (zti |xt ) are not cond. independent
Sensor models can only be
... ... n
1 2 n
zt−1 1 2 zt
zt−1 zt−1 zt zt generated seperatly for each beam
Assumption that measurements
are independent: “Works
i
P (zt |xt) ˆi
zt zmax
surprisingly well”. . . if. . .
increasing uncertainty (tweaking)
using every 10th measurement
zrand ...
13. O
1 I: S E U D M
2 T D M CRF
Short Introduction to CRF
CRF-Model for State Estimation
3 A
CRF-Filter Algorithm
Learning the Parameters
4 E R
14. O
1 I: S E U D M
2 T D M CRF
Short Introduction to CRF
CRF-Model for State Estimation
3 A
CRF-Filter Algorithm
Learning the Parameters
4 E R
15. Intro Transformation of Directed Model to CRF Application Experimental Results
I: CRF
Undirected graphical models
ut−2 ut−1
xt−2 xt−1 xt
zt−1 zt
16. Intro Transformation of Directed Model to CRF Application Experimental Results
I: CRF
Undirected graphical models
Every (possible) dependency
ut−2 ut−1 represented by edge
xt−2 xt−1 xt
zt−1 zt
17. Intro Transformation of Directed Model to CRF Application Experimental Results
I: CRF
Undirected graphical models
Every (possible) dependency
ut−2 ut−1 represented by edge
Distribution defined over products
xt−2 xt−1 xt
of functions over cliques
zt−1 zt
18. Intro Transformation of Directed Model to CRF Application Experimental Results
I: CRF
Undirected graphical models
Every (possible) dependency
ut−2 ut−1 represented by edge
Distribution defined over products
xt−2 xt−1 xt
of functions over cliques
zt−1 zt Functions are called clique
potentials
19. Intro Transformation of Directed Model to CRF Application Experimental Results
I: CRF
Undirected graphical models
Every (possible) dependency
ut−2 ut−1 represented by edge
Distribution defined over products
xt−2 xt−1 xt
of functions over cliques
zt−1 zt Functions are called clique
potentials
Clique potentials represent
compatibility of their variables
20. O
1 I: S E U D M
2 T D M CRF
Short Introduction to CRF
CRF-Model for State Estimation
3 A
CRF-Filter Algorithm
Learning the Parameters
4 E R
21. Intro Transformation of Directed Model to CRF Application Experimental Results
CRF-M S E
ut−2 ut−1
xt−2 xt−1 xt
zt−1 zt
T
1
p (x0:T |z1:T , u0:T −1 ) = ϕp (xt , xt −1 , ut −1 )ϕm (xt , zt )
Z (z1:T , u1:T −1 )
t =1
22. Intro Transformation of Directed Model to CRF Application Experimental Results
CRF-M S E
ut−2 ut−1
xt−2 xt−1 xt
zt−1 zt
T
1
p (x0:T |z1:T , u0:T −1 ) = ϕp (xt , xt −1 , ut −1 )ϕm (xt , zt )
Z (z1:T , u1:T −1 )
t =1
Z (·): all trajectories ϕp (·)ϕm (·)
23. Intro Transformation of Directed Model to CRF Application Experimental Results
CRF-M S E
ut−2 ut−1
xt−2 xt−1 xt
zt−1 zt
T
1
p (x0:T |z1:T , u0:T −1 ) = ϕp (xt , xt −1 , ut −1 )ϕm (xt , zt )
Z (z1:T , u1:T −1 )
t =1
Z (·): all trajectories ϕp (·)ϕm (·)
How to define ϕp (·) and ϕm (·)?
24. Intro Transformation of Directed Model to CRF Application Experimental Results
T P P φp
ut −1 = (δrot1 , δtrans , δrot2 ) odometry
ut −1 = (δrot1 , δtrans , δrot2 ) derived odometry
ˆ ˆ ˆ ˆ
δrot2
2
Before: Gaussian noise N uti −1 , σi
xt
δtrans
δrot1
xt−1
27. Intro Transformation of Directed Model to CRF Application Experimental Results
T P P φp
ut −1 = (δrot1 , δtrans , δrot2 ) odometry
ut −1 = (δrot1 , δtrans , δrot2 ) derived odometry
ˆ ˆ ˆ ˆ
δrot2
2
Before: Gaussian noise N uti −1 , σi
xt
(δrot1 − δrot1 )2
ˆ
δtrans
fp (xt , xt −1 , ut −1 ) = (δtrans − δtrans )2
ˆ
3 features
(δrot2 − δrot2 )2
ˆ
δrot1 φp (xt , xt −1 , ut −1 ) = exp wp , fp (xt , xt −1 , ut −1 )
xt−1 1 (a − a )2
ˆ
N a, = exp −
σ2 2σ2
Gaussian noise N uti −1 , 1
−2wpi if wp < 0
i
28. Intro Transformation of Directed Model to CRF Application Experimental Results
R: S M N¨ B A
i
P (zt |xt) ˆi
zt zmax
zrand
n
p (zt |xt ) = p (zti |xt )
i =1
36. O
1 I: S E U D M
2 T D M CRF
Short Introduction to CRF
CRF-Model for State Estimation
3 A
CRF-Filter Algorithm
Learning the Parameters
4 E R
37. O
1 I: S E U D M
2 T D M CRF
Short Introduction to CRF
CRF-Model for State Estimation
3 A
CRF-Filter Algorithm
Learning the Parameters
4 E R
38. Intro Transformation of Directed Model to CRF Application Experimental Results
U CRF P F
At each time step t:
Prediction
Move particles according to gaussian noise
determined by wp
Same as sampling from N uti −1 ,
ˆ 1
i
−2wp
Correction
Particle at xt gets weight φm (xt , zt )
Resample (includes normalization)
39. Intro Transformation of Directed Model to CRF Application Experimental Results
U CRF P F
At each time step t:
Prediction
Move particles according to gaussian noise
determined by wp
u
Same as sampling from N uti −1 ,
ˆ 1
i
−2wp
Correction
Particle at xt gets weight φm (xt , zt )
Resample (includes normalization)
40. Intro Transformation of Directed Model to CRF Application Experimental Results
U CRF P F
At each time step t:
Prediction
Move particles according to gaussian noise
determined by wp moved
Same as sampling from N uti −1 ,
ˆ 1
−2wpi particles
Correction
Particle at xt gets weight φm (xt , zt )
Resample (includes normalization)
41. Intro Transformation of Directed Model to CRF Application Experimental Results
U CRF P F
At each time step t:
Prediction
Move particles according to gaussian noise
determined by wp added
Same as sampling from N uti −1 , noise
1
ˆ −2wpi
Correction
Particle at xt gets weight φm (xt , zt )
Resample (includes normalization)
42. Intro Transformation of Directed Model to CRF Application Experimental Results
U CRF P F
At each time step t:
Prediction
Move particles according to gaussian noise
determined by wp ...sense...
Same as sampling from N uti −1 ,
ˆ 1
−2wpi
Correction
Particle at xt gets weight φm (xt , zt )
Resample (includes normalization)
43. Intro Transformation of Directed Model to CRF Application Experimental Results
U CRF P F
At each time step t:
Prediction
Move particles according to gaussian noise
determined by wp weights
Same as sampling from N uti −1 ,
ˆ 1
−2wpi
Correction
Particle at xt gets weight φm (xt , zt )
Resample (includes normalization)
44. Intro Transformation of Directed Model to CRF Application Experimental Results
U CRF P F
At each time step t:
Prediction
Move particles according to gaussian noise
determined by wp
resample
Same as sampling from N uti −1 ,
ˆ 1
i
−2wp
Correction
Particle at xt gets weight φm (xt , zt )
Resample (includes normalization)
45. O
1 I: S E U D M
2 T D M CRF
Short Introduction to CRF
CRF-Model for State Estimation
3 A
CRF-Filter Algorithm
Learning the Parameters
4 E R
46. Intro Transformation of Directed Model to CRF Application Experimental Results
D wp wm
Drive around in test area
47. Intro Transformation of Directed Model to CRF Application Experimental Results
D wp wm
Drive around in test area
Use high-quality scanmatcher to generate
“ground truth” trajectory x∗
48. Intro Transformation of Directed Model to CRF Application Experimental Results
D wp wm
Drive around in test area
Use high-quality scanmatcher to generate
“ground truth” trajectory x∗
ˆ
Using arbitrary weights, generate trajectory x
with CRF-filter
49. Intro Transformation of Directed Model to CRF Application Experimental Results
D wp wm
Drive around in test area
Use high-quality scanmatcher to generate
“ground truth” trajectory x∗
ˆ
Using arbitrary weights, generate trajectory x
with CRF-filter
Use difference of summed features as weight
update(−) :
wk = wk −1 + α ( f (x∗ , u, z) − f (x, u, z))
ˆ
50. Intro Transformation of Directed Model to CRF Application Experimental Results
D wp wm
Drive around in test area
Use high-quality scanmatcher to generate
“ground truth” trajectory x∗
ˆ
Using arbitrary weights, generate trajectory x
with CRF-filter
Use difference of summed features as weight
update(−) :
wk = wk −1 + α ( f (x∗ , u, z) − f (x, u, z))
ˆ
Decrease α if new Filter cannot track
51. Intro Transformation of Directed Model to CRF Application Experimental Results
D wp wm
Drive around in test area
Use high-quality scanmatcher to generate
“ground truth” trajectory x∗
ˆ
Using arbitrary weights, generate trajectory x
with CRF-filter
Use difference of summed features as weight
update(−) :
wk = wk −1 + α ( f (x∗ , u, z) − f (x, u, z))
ˆ
Decrease α if new Filter cannot track
loop
Adapts weights to task, sensor dependencies/environment,
sensor noise, particle filter parameters
52. Intro Transformation of Directed Model to CRF Application Experimental Results
L A
Averaged Perceptron Algorithm (Collins 2002) for tagging
w k = w k −1 + α f (x∗ , u, z) − f (x, u, z)
ˆ
Proven to converge even in presence of errors in training data
Intuition of learning algorithm:
If PF works correctly, then
f (xn , un−1 , zn ) =
∗
f (xn , un−1 , zn )
ˆ
f i occurs less often in x∗ than in x → decrease influence of f i
ˆ
on particle filter by decreasing w i
53. O
1 I: S E U D M
2 T D M CRF
Short Introduction to CRF
CRF-Model for State Estimation
3 A
CRF-Filter Algorithm
Learning the Parameters
4 E R
54. Intro Transformation of Directed Model to CRF Application Experimental Results
E R
Properties of the learned weights
Norm of weight vector decreases with
number of laser beams in z
believes the features/measurements less
equivalent to initially introduced
“tweaking”?!
55. Intro Transformation of Directed Model to CRF Application Experimental Results
E R
Properties of the learned weights
Norm of weight vector decreases with
number of laser beams in z
believes the features/measurements less
equivalent to initially introduced
“tweaking”?!
Two specialized CRF-filters compared to
generative particle filter trained using
expectation maximization
Tracking Global
Error Localization
Accuracy
Generative 7.52 cm 30%
CRF-Filter 7.07 cm 96%
56. Intro Transformation of Directed Model to CRF Application Experimental Results
C
1 A CRF is an alternative, undirected graphical model
57. Intro Transformation of Directed Model to CRF Application Experimental Results
C
1 A CRF is an alternative, undirected graphical model
2 CRF-Filters use a continuous CRF for recursive state
estimation
58. Intro Transformation of Directed Model to CRF Application Experimental Results
C
1 A CRF is an alternative, undirected graphical model
2 CRF-Filters use a continuous CRF for recursive state
estimation
3 . . . can be trained to maximize filter performance depending
on the task
59. Intro Transformation of Directed Model to CRF Application Experimental Results
C
1 A CRF is an alternative, undirected graphical model
2 CRF-Filters use a continuous CRF for recursive state
estimation
3 . . . can be trained to maximize filter performance depending
on the task
4 . . . can deal with correlated measurements
60. Intro Transformation of Directed Model to CRF Application Experimental Results
C
1 A CRF is an alternative, undirected graphical model
2 CRF-Filters use a continuous CRF for recursive state
estimation
3 . . . can be trained to maximize filter performance depending
on the task
4 . . . can deal with correlated measurements
5 . . . do not explicitly account for dependencies between sensor
data