HAL Id: hal-02468625
https://hal.inria.fr/hal-02468625
Submitted on 6 Feb 2020
HAL is a multi-disciplinary open access
archive for the deposit and dissemination of
sci-entific research documents, whether they are
pub-lished or not. The documents may come from
teaching and research institutions in France or
abroad, or from public or private research centers.
L’archive ouverte pluridisciplinaire HAL, est
destinée au dépôt et à la diffusion de documents
scientifiques de niveau recherche, publiés ou non,
émanant des établissements d’enseignement et de
recherche français ou étrangers, des laboratoires
publics ou privés.
Topological navigation using sensory memory: SLAN
versus SLAM
Philippe Martinet, J. Courbon, Y. Mezouar, B. Thuilot
To cite this version:
Philippe Martinet, J. Courbon, Y. Mezouar, B. Thuilot. Topological navigation using sensory
mem-ory: SLAN versus SLAM. Journée SLAM, Dassault-Aviation, Dec 2009, Saint-Cloud, France.
�hal-02468625�
Philippe Martinet
Topological navigation using sensory memory: SLAN versus SLAM
LASMEA, Blaise Pascal University, Clermont-Ferrand
Topological navigation using sensory memory
SLAN versus SLAM
Professor Philippe Martinet
J. Courbon, Y. Mezouar, B. Thuilot
LASMEA Laboratory
Blaise Pascal University
martinet@lasmea.univ-bpclermont.fr
http://wwwlasmea.univ-bpclermont.fr/Control
http://robots.lasmea.univ-bpclermont.fr
Dassault Aviation, Saint-Cloud, France, December 16
th
2009
Philippe Martinet
Topological navigation using sensory memory: SLAN versus SLAM
LASMEA, Blaise Pascal University, Clermont-Ferrand
Program of the journey
2
LASMEA
P. Martinet
Navigation toplogique par mémoire sensorielle: SLAN
vs SLAM
12H
IGN/MATIS
N. Paparoditis
16H
Inria-Sophia
Antipolis
P. Rives
Outils et méthodes d'estimation pour le SLAM
-Application au Slam visuel
14H
ENSTA
ParisTech/UEI Lab
D. Filliat
Détection visuelle de fermeture de boucle et applications
au SLAM métrique et topologique
15H
Dassault Aviation
B. Patin
Cloture du séminaire
17H
LAAS
Juan Sola
Initialisation immédiate d-amers en SLAM monoculaire.
Points et lignes
11H
Inria-Sophia
Antipolis
P. Rives
Les enjeux et problématiques du SLAM (Survey)
10H15
Dassault Aviation
B. Patin
Accueil et présentation de la journée
10H
Philippe Martinet
Topological navigation using sensory memory: SLAN versus SLAM
LASMEA, Blaise Pascal University, Clermont-Ferrand
Outline of the presentation
Introduction
•
LASMEA
•
Sensor based Control
•
Navigation
•
SLAN vs SLAM
Togological navigation system
•
Global system
•
Case of vision
•
Differents Modules
•
SOVIN
Discussion
•
Learning process
•
Updating process
•
Closing the loop problem
A generic model
•
A unified model for camera
•
Case of fisheye cameras
•
Partial euclidian reconstruction
Illustrations
•
Indoor applications
•
Outdoor applications
3 Philippe
Martinet
Topological navigation using sensory memory: SLAN versus SLAM
LASMEA, Blaise Pascal University, Clermont-Ferrand
Introduction
A generic
model
Topological
navigation system
Illustrations
Discussion
Lasmea
400km south of Paris
200km West of Lyon
Mixed Unit 6602
CNRS/UBP
LASMEA
Universitary Campus
France
Clermont-Ferrand
Director : Michel Dhome
4
Philippe Martinet
Topological navigation using sensory memory: SLAN versus SLAM
LASMEA, Blaise Pascal University, Clermont-Ferrand
Introduction
A generic
model
Topological
navigation system
Illustrations
Discussion
Lasmea
400km south of Paris
200km West of Lyon
Mixed Unit 6602
CNRS/UBP
Director : Michel Dhome
MATELEC
¾Optoelectronics,
microelectronics
¾Electromagnetism
¾Gaz sensor
GRAVIR
¾Perception System
¾Computer that See
¾ROSACE
33 teacher/researcher
1 research scientist
25 Phd
25 teacher/researcher
2 research scientist,
49 Phd, 2 post-doc
5 Philippe MartinetTopological navigation using sensory memory: SLAN versus SLAM
LASMEA, Blaise Pascal University, Clermont-Ferrand
Introduction
A generic
model
Topological
navigation system
Illustrations
Discussion
6
Lasmea : GRAVIR-PERSYST
J.P. Derutin
Perception System
PERception SYSTems
Multisensor
Perception
Data fusion
Scene understanding
SLAM
Smart Sensors
Algorithm Architecture Adequation
Cognitive Vision
Architecture and
methods
Parallelism in vision
High speed prototyping tools
Tools for embedded applications
Philippe Martinet
Topological navigation using sensory memory: SLAN versus SLAM
LASMEA, Blaise Pascal University, Clermont-Ferrand
Introduction
A generic
model
Topological
navigation system
Illustrations
Discussion
Lasmea : GRAVIR-COMSEE
T. Chateau
Artificial Vision
COMputer that SEE
Geometry for visual
perception
Metrology by vision
Automatic rigid scene reconstruction
Vision in Deformable
Environments
Modeling and algorithm for deformable
environment
Visual
Recognition
Real time object tracking
Object recognition
http://comsee.univ-bpclermont.fr
7 Philippe
Martinet
Topological navigation using sensory memory: SLAN versus SLAM
LASMEA, Blaise Pascal University, Clermont-Ferrand
Introduction
A generic
model
Topological
navigation system
Illustrations
Discussion
8
Lasmea : GRAVIR-ROSACE
P. Martinet & Y. Mezouar
ROSACE
ROboticS and Autonomous Complex systEms
VISIR
VIsual ServoIng of Robots
Omnidirectional visual servoing
Topological navigation through sensory memory
Multi-sensor-based control
AGV
Automatic Guided Vehicles
Control design under uncertain dynamics
Hybrid control architecture
Multi-robot-system control
MICMAC
Modeling, Identification and
Control of Complex Machines
Vision-Based Control of Parallel Robots
Control of redundant robots
Control of High Dynamics System
http://wwwlasmea.univ-bpclermont.fr/Control http://robots.lasmea.univ-bpclermont.fr400Hz
1kHz
…
15 to 60Hz
…
15 to 60Hz
Philippe MartinetTopological navigation using sensory memory: SLAN versus SLAM
LASMEA, Blaise Pascal University, Clermont-Ferrand
Introduction
A generic
model
Topological
navigation system
Illustrations
Discussion
9
Lasmea : GRAVIR
People involved in SLAM or near SLAM problems
PERSYST
F. Berry, R. Chapuis, F. Chausse, P. Checchin, J.P. Derutin, L.
Trassoudaine
(IMPALA/CITYVIP)
(CITYHOME/EURIPIDE/BRI)
COMSEE
M. Dhome, M. Lhuillier , E. Royer
(MOBIVIP/CITYVIP/BODEGA)
(FACT/CITYHOME/EURIPIDE/BRI)
ROSACE
Y. Mezouar, P. Martinet, B. Thuilot
(MOBIVIP/CITYVIP/BODEGA/R-DISCOVER/ARMEN)
(SAFEMPOVE, FACT/CITYHOME/EURIPIDE/BRI/)
Autonomous navigation (RTK-GPS, Vision, Multi-sensor)
Platoon: Autonomous leader (RTK-GPS, Vision, Multi-sensor)
Platoon: Leader manually driven (RTK-GPS, Vision)
2D/3D maps building (vision, Radar, Range finder, Multi-sensor)
Applications
Philippe Martinet
Topological navigation using sensory memory: SLAN versus SLAM
LASMEA, Blaise Pascal University, Clermont-Ferrand
Introduction
A generic
model
Topological
navigation system
Illustrations
Discussion
10
Sensor based control: History
˜
y
˜θ
M
= (
s, C
(s)
)
d
0d
1d
2d
5d
4d
3Lateral control in cartesian space:
Lateral control in sensor space:
Bringing
³
y, ˜
˜
θ
´
to (0, 0)
∆
Explicite declaration of ∆ and
required cartesian localization
d =(d
0, d
1, d
2, d
3, d
4, d
5)
TBringing d to d
∗No explicite declaration of ∆ and
do not required cartesian localization
d
∗3d
∗ 4d
∗5d
∗0d
∗ 1d
∗ 2 [ESPIAU87]Philippe Topological navigation using sensory memory: SLAN versus SLAM
Introduction
A generic
model
Topological
navigation system
Illustrations
Discussion
11
Sensor based control: Basic concepts
x
y
z
Fo
x
y
z
Fc
x
y
z
Fa
x
y
z
Fc
Case of embedded cameraω
υ
x
: 3D pose
If
(fixed object)∂s∂t= 0
˙s = L
s
v
Interaction matrix
˙s = L
s
v
+
∂s
∂t
s
= s(x, t)
˙s =
∂s
∂x
dx
dt
+
∂s
∂t
˙s =
∂s
∂x
v
+
∂s
∂t
[SAMSON90,ESPIAU92]Philippe Topological navigation using sensory memory: SLAN versus SLAM
Introduction
A generic
model
Topological
navigation system
Illustrations
Discussion
12
Navigation
Planning
Mission, route
Localization
Single, platoon modeObstacle
avoidance
Scheduling
Execution level
Control
Single, platoon modeFDIR
monitoring
End user level
Reference trajectory
Modified reference trajectory
Philippe Martinet
Topological navigation using sensory memory: SLAN versus SLAM
LASMEA, Blaise Pascal University, Clermont-Ferrand
Introduction
model
Topological
navigation system
Illustrations
Discussion
13
Navigation
Navigation scheme
Planning
Mission, route
Obstacle
avoidance
Control
Single, platoon modeLocalization
Single, platoon mode
Reference trajectory
Scheduling
Execution level
FDIR
monitoring
End user level
Philippe Martinet
Topological navigation using sensory memory: SLAN versus SLAM
LASMEA, Blaise Pascal University, Clermont-Ferrand
Introduction
model
Topological
navigation system
Illustrations
Discussion
14
Navigation
3D Map based navigation (absolute navigation)
Online step
Exteroceptive
sensors
Proprioceptive
sensors
Itinerary
Selection
& Execution
Localization
GIS 3D map based Reference trajectory Global Lateral deviation Angular deviation Curvature Curvilinear abscissa Philippe MartinetTopological navigation using sensory memory: SLAN versus SLAM
LASMEA, Blaise Pascal University, Clermont-Ferrand
Introduction
A generic
model
Topological
navigation system
Illustrations
Discussion
15
Navigation
Learning step
Exteroceptive
sensors
Proprioceptive
sensors
GIS
Itinerary
Selection
& Execution
Topological
Indexation
Extraction
of features
Augmented GIS
Topological
Representation
sensory memory
Memory based navigation (relative navigation)
Philippe Martinet
Topological navigation using sensory memory: SLAN versus SLAM
LASMEA, Blaise Pascal University, Clermont-Ferrand
Introduction
A generic
model
Topological
navigation system
Illustrations
Discussion
16
Navigation
Memory based navigation (relative navigation)
Online step
Exteroceptive
sensors
Proprioceptive
sensors
Itinerary
Selection
& Execution
Localization
Augmented GIS Topological Representation Reference trajectory LOCAL Lateral deviation Angular deviation Curvature Curvilinear abscissa Philippe MartinetTopological navigation using sensory memory: SLAN versus SLAM
LASMEA, Blaise Pascal University, Clermont-Ferrand
Introduction
A generic
model
Topological
navigation system
Illustrations
Discussion
17
SLAN vs SLAM
Sensor space
Cartesian (relative)
Cartesian
Control space
Relative
Absolute
Localization
Topological
Semi-Metric
Metric 2D/3D
MAPS
SLAN
SLAM
Philippe MartinetTopological navigation using sensory memory: SLAN versus SLAM
LASMEA, Blaise Pascal University, Clermont-Ferrand
Outline of the presentation
Introduction
•
LASMEA
•
Sensor based Control
•
Navigation
•
SLAN vs SLAM
Togological navigation system
•
Global system
•
Case of vision
•
Differents Modules
•
SOVIN
Discussion
•
Learning process
•
Updating process
•
Closing the loop problem
A generic model
•
A unified model for camera
•
Case of fisheye cameras
•
Partial euclidian reconstruction
Illustrations
•
Indoor applications
•
Outdoor applications
Philippe Martinet
Topological navigation using sensory memory: SLAN versus SLAM
LASMEA, Blaise Pascal University, Clermont-Ferrand
Introduction
A generic
model
Topological
navigation system
Illustrations
Discussion
19
A unified model for camera
Single view point system :
Classification
F’ F F F’
Planar
Elliptical
FParabolic
F’ FHyperbolic
Log/polar retina[Geyer00]
Philippe MartinetTopological navigation using sensory memory: SLAN versus SLAM
LASMEA, Blaise Pascal University, Clermont-Ferrand
Introduction
A generic
model
Topological
navigation system
Illustrations
Discussion
20
A unified model for camera
F’ F F’
Planar
Elliptical
FParabolic
F’ FHyperbolic
Mirror type
Parabolic
1
1+2p
Hyperbolic
Elliptical
Planar
0
1
Conventional
0
1
ϕ
ϕ and ξ : Mirror parameters
ξ
d+2p
√
d
2+4p
2d
−2p
√
d
2+4p
2d
√
d
2+4p
2d
√
d
2+4p
2Single view point system :
Classification
[Geyer00]
Philippe Martinet
Topological navigation using sensory memory: SLAN versus SLAM
LASMEA, Blaise Pascal University, Clermont-Ferrand
Introduction
A generic
model
Topological
navigation system
Illustrations
Discussion
21
A unified model for camera
m
F
c
F
Mirror : unitary sphere
ξ
m3D(X, Y, Z)Image plane
m
pm
nK
Mm
mm
m=
1ρ⎡
⎣
X
Y
Z
⎤
⎦
ρ =
kmk =
√
X
2+ Y
2+ Z
2m
p
= K
M
m
n
m
n= [x
Tβ]
T=
⎡
⎢
⎢
⎢
⎣
X
Z + ξρ
Y
Z + ξρ
β
⎤
⎥
⎥
⎥
⎦
Single view point system :
case of point
Philippe Martinet
Topological navigation using sensory memory: SLAN versus SLAM
LASMEA, Blaise Pascal University, Clermont-Ferrand
Introduction
A generic
model
Topological
navigation system
Illustrations
Discussion
22
A unified model for camera
Generic Projection function
m F c F
mirror
Image plane
Special cases :
Î perspective projection
Î spherical projection
ϕ and ξ : Mirror parameters
ϕ - 2 ξ
ξ
m
pM
=
⎡
⎣
ϕ
− ξ
0
ϕ
− ξ 0
0
0
0
0
1
⎤
⎦
K
=
⎡
⎣
f
0
f s
f r
u
v
0
0
0
0
1
⎤
⎦
m3D(X, Y, Z)m
p= K M
| {z }
f(m
3D)
K
Mm
m² = 1 and ξ = 0
² = 0 and ξ = 1
f(m
3D) =
⎡
⎢
⎣
X ²Z+ξ√X2+Y2+Z2 Y ²Z+ξ√X2+Y2+Z21
⎤
⎥
⎦
Single view point system :
case of point
Philippe Topological navigation using sensory memory: SLAN versus SLAM
Introduction
A generic
model
Topological
navigation system
Illustrations
Discussion
23
Case of fisheye camera
Fisheye camera model
:
definition
Pinhole camera
Fisheye camera
F
r = distance between X and the image point
F
θ = angle between the incoming ray and the principal axis
F
Fm
: camera frame
F
Fi
: image frame
Philippe Topological navigation using sensory memory: SLAN versus SLAM
Introduction
A generic
model
Topological
navigation system
Illustrations
Discussion
24
Case of fisheye camera
Fisheye camera model
:
Pinhole based models
one parameter
n parameters
F
r
1
f
(r
p
) = r
p
L(r
p
, n)
F
r
2
f
(r
p
) =
L(r
r
p p,n)
F
r
f
3
(r
p
) =
r
p1+k
1r
2pMapping :
[T. Pajdla97,Hartley00,Zhang98,Ma2006…]
[Fitzgibbon01]
[Fitzgibbon01]
Philippe Martinet
Topological navigation using sensory memory: SLAN versus SLAM
LASMEA, Blaise Pascal University, Clermont-Ferrand
Introduction
model
Topological
navigation system
Illustrations
Discussion
25
Case of fisheye camera
n1+n2 parameters
F
r
4
f
(r
p
) = r
p
L
L
1(r
p,n
1)
2(r
p,n
2)
F
r
5
f
(r
p
) = s log(1 + λr
p
)
F
r
f
(r
p
) = λ
r
pF
r
6
f
(r
p
) =
ω
1
arctan
¡
2r
p
tan
ω
2
¢
s is a scaling factor and λ a gain to control the amount of distortion
The ”distortion function” λ =
{λ1
, λ
2, . . . λ
n} can be then fitted by apara-metric model
[Hartley07]
[Li05]
[Basu95]
[Devernay01]
Fisheye camera model
:
Pinhole based models
Mapping :
Philippe Martinet
Topological navigation using sensory memory: SLAN versus SLAM
LASMEA, Blaise Pascal University, Clermont-Ferrand
Introduction
model
Topological
navigation system
Illustrations
Discussion
26
Case of fisheye camera
Fisheye camera model
:
Captured rays based models
Mapping :
F
r
1
f
(θ) = f θ
F
r
3
f
(θ) = f sin θ
F
r
2
f
(θ) = 2f tan
¡
θ
2
¢
F
r
4
f
(θ) = f sin
¡
θ
2
¢
cameras with limited distortions
[Kingslake89]
stereographic projection
preserves circularity and thus project 3D
local symmetries onto 2D local symmetries
[Fleck94,Stevenson95]
orthogonal or sine law projection
[Ray94]
equisolid angle projection
[Smith92]
Philippe Martinet
Topological navigation using sensory memory: SLAN versus SLAM
LASMEA, Blaise Pascal University, Clermont-Ferrand
Introduction
A generic
model
Topological
navigation system
Illustrations
Discussion
27
Case of fisheye camera
Fisheye camera model
:
Captured rays based models
Mapping :
F
r
5
f
(θ) = f (k
1
θ + k
2
θ
3
+
· · · + k
n
θ
2(n
−1)+1
)
F
r
6
f
(θ) = α sin(βθ)
F
r
7
f
(θ) = a tan(θ/b) + c sin(θ/d)
Improvment of the polynomial model accuracy
[xiong97,kannala04,schwalbe05,Scaramuzza06]
[Kumler00]
F
α : scale factor
F
β : radial mapping parameter
[Bakstein02]
Combination of
- stereographic projection (with parameters a, b)
- equisolid angle projection (with parameters c, d)
Philippe Martinet
Topological navigation using sensory memory: SLAN versus SLAM
LASMEA, Blaise Pascal University, Clermont-Ferrand
Introduction
A generic
model
Topological
navigation system
Illustrations
Discussion
28
Case of fisheye camera
F
F
c: frame attached to the conventional camera
F
Fm
: frame attached to the unitary sphere
Optical
center
Principal
Projection
center
Unitary
sphere
X = [X Y Z]
Tx = [x y 1]
TStep 1 : Projection on the unitary
sphere
X
m
in F
m: X
m= X/ρ
where ρ =
kXk =
√
X
2+ Y
2+ Z
2Step 2 : Perspective projection
on the normalized image plane Z = 1
− ξ
x’=f(X) =
∙
X
εsZ + ξρ
Y
εsZ + ξρ
1
¸>
spherical projection : ε
s= 0 and ξ = 1
Catadioptric model : εs
= 1 and ξ = 1
pinhole model : ε
s= 1 and ξ = 0
Generic camera model
:
Unified Spherical Model
Philippe Martinet
Topological navigation using sensory memory: SLAN versus SLAM
LASMEA, Blaise Pascal University, Clermont-Ferrand
Introduction
A generic
model
Topological
navigation system
Illustrations
Discussion
29
Case of fisheye camera
Generic camera model
:
Unified Spherical Model
F
Fc
: frame attached to the conventional camera
F
F
m: frame attached to the unitary sphere
Optical
center
Principal
Projection
center
Unitary
sphere
X = [X Y Z]
Tx = [x y 1]
TStep 3 : Projection onto image plane
x = M x’
M =
⎛
⎝
f κ
0
0
0
δf κ
0
0
0
1
⎞
⎠ δ is equal to ±1
κ > 0
f : focal length
perspective camera : κ = 1 and δ = +1
catadioptric cameras : δ =
−1
Philippe Martinet
Topological navigation using sensory memory: SLAN versus SLAM
LASMEA, Blaise Pascal University, Clermont-Ferrand
Introduction
A generic
model
Topological
navigation system
Illustrations
Discussion
30
Case of fisheye camera
Generic camera model
:
Unified Spherical Model
F
Fc
: frame attached to the conventional camera
F
F
m: frame attached to the unitary sphere
Optical
center
Principal
Projection
center
Unitary
sphere
X = [X Y Z]
Tx = [x y 1]
TStep 4 : Pixel transformation
m = K x
plane-to-plane collineation K
K =
⎛
⎝
k
us
uvu
00
k
vv
00
0
1
⎞
⎠
(u0, v0)
T: position of the optical center
k
uand k
v: scaling along the x and y axes
s
uv: skew
setting ξ = 0, κ = 1 and δ = +1
½
x
p= f
pX/Z
y
p= f
pY /Z
Pinhole camera
Philippe Martinet
Topological navigation using sensory memory: SLAN versus SLAM
LASMEA, Blaise Pascal University, Clermont-Ferrand
Introduction
A generic
model
Topological
navigation system
Illustrations
Discussion
31
Case of fisheye camera
Generic camera model
:
Unified Spherical Model
For fisheye camera
r
f= r
f(θ) =
1+ξf√ftantan θ2θ+1is a T
2-mapping linking the radius r
fand the incidence angle θ
r
f= r
f(r
p) =
ff fprp 1+ξ s r2p f 2p+1Constraint 1 : easily verified (r
f(0) = 0)
Constraint 2 : r
f(k) is monotonically increasing for k > 0
is a T
1-mapping linking the perspective Radius
rp
and the fisheye radius rf
The generic camera model is candidate
for fisheye camera modeling
Philippe Martinet
Topological navigation using sensory memory: SLAN versus SLAM
LASMEA, Blaise Pascal University, Clermont-Ferrand
Introduction
A generic
model
Topological
navigation system
Illustrations
Discussion
32
Case of fisheye camera
Validation with calibration toolbox [MEI07]
Computation of f
f
and ξ (Producers data)
Model fitting for all existing fisheye cameras
Calibration
List of fisheye cameras considered
List of models considered
Polynomial model using perspective projection
Polynomial model using unified model projection
Proposed unified model
r
f(r
p) = r
p(1 + a
1r
2p+
· · · + a3
r
6p)
r
f(r
u) = r
u(1 + a
1r
2u+
· · · + a3
r
u6)
r
f= r
uThe generic camera model is validated with an average
of 0.2 pixels reprojection error for fisheye camera modeling
Only one parameter required
Philippe Martinet
Topological navigation using sensory memory: SLAN versus SLAM
LASMEA, Blaise Pascal University, Clermont-Ferrand
Introduction
A generic
model
Topological
navigation system
Illustrations
Discussion
33
Partial euclidian reconstruction
> 0
Introducing
where
γ
x
=
p
1 + (1
− ξ
2
)(x
2
+ y
2
)
with
m
n
Unified model - Case of points
η = s
ρ |Z|=
p
1 + X
2/Z
2+ Y
2/Z
2 s=sign(Z)
x =
1+ξηX/Zy =
1+ξηY /Z(1 + ξη) x =
X Z(1 + ξη) y =
Y Zη
2− (x
2+ y
2)(1 + ξη)
2− 1 = 0
η =
±γ
x
+ξ(x
2
+y
2
)
1−ξ
2
(x
2
+y
2
)
η computed without ambiguity
m
m
= (η
−1
+ ξ)m
m
= [x y
1+ξη
1
]
T
Philippe Martinet
Topological navigation using sensory memory: SLAN versus SLAM
LASMEA, Blaise Pascal University, Clermont-Ferrand
Introduction
A generic
model
Topological
navigation system
Illustrations
Discussion
34
Partial euclidian reconstruction
mirror mirror
m
mF
mF
cm
∗ mF
∗ mF
∗ cm
nm
∗ nm
π
(R,t)
Reference plane π in F
∗ m[X Y Z]
Tm
coordinates in
F
m[X
∗Y
∗Z
∗]
Tm
coordinates in
F
∗ md
∗n
∗π
∗T= [n
∗T− d
∗]
Unified model - Case of points
Philippe Topological navigation using sensory memory: SLAN versus SLAM
Introduction
A generic
model
Topological
navigation system
Illustrations
Discussion
35
Partial euclidian reconstruction
Homogeneous coordinates:
m
= [X Y Z H]
T
and m
∗
= [X
∗
Y
∗
Z
∗
H
∗
]
T
The distance from the world point m to the plane (π)
m
m= (η
−1+ ξ)m =
1ρ£
X
Y
Z
¤T
m
∗ m= (η
∗−1+ ξ)m
∗=
ρ1∗£
X
∗Y
∗Z
∗¤T
m
= [x y
1+ξη
1
]
T
m
∗
= [x
∗
y
∗
1
1+ξη
∗]
T
ρ(η
−1+ ξ)m =
£
I
30
¤
m
=
£
R
t
¤
m
∗d(m, π) = π
∗T· m
∗= [n
∗T− d
∗][X
∗Y
∗Z
∗H
∗]
T= ρ
∗(η
∗−1+ ξ)n
∗Tm
∗−
d
∗H
∗Unified model - Case of points
Philippe Topological navigation using sensory memory: SLAN versus SLAM
Introduction
A generic
model
Topological
navigation system
Illustrations
Discussion
36
Partial euclidian reconstruction
where
b
∗ π=
h
0
1×3−
d(m,π)d∗i
A
∗ π=
h
I
3 n ∗ d∗i
TEuclidian Homography
with H
π
= R +
d
t
∗n
∗T
and α =
−
d(m,π)
d
∗H
∗
=
ρ
∗(η
∗−1+ξ)
d
∗n
∗T
m
∗
−
d(m,π)
d
∗ρ
∗(η
∗−1+ ξ)m
∗=
£
X
∗Y
∗Z
∗¤T
m
∗
= ρ
∗
(η
∗−1
+ ξ)A
∗
π
m
∗
+ b
∗
π
ρ(η
−1+ ξ)m =
£
R
t
¤
m
∗ρ(η
−1
+ ξ)m = ρ
∗
(η
∗−1
+ ξ)H
π
m
∗
+ αt
Philippe Martinet
Topological navigation using sensory memory: SLAN versus SLAM
LASMEA, Blaise Pascal University, Clermont-Ferrand
Introduction
model
Topological
navigation system
Illustrations
Discussion
37
Partial euclidian reconstruction
Unified model - Case of points
H
πR
and t
d∗=
dt∗m
= β
x,x
∗
H
π
m
∗
Linear form
m
× H
πm
∗= 0
where β
x,x
∗=
η
∗−1+ξ
η
−1+ξ
ρ
∗ρ
σ =
ρ
ρ
∗= (1 + n
∗T
R
T
t
d
∗)
(η
∗−1+ξ)n
∗Tm
∗(η
−1+ξ)n
∗TR
Tm
m
∈ π
Philippe MartinetTopological navigation using sensory memory: SLAN versus SLAM
LASMEA, Blaise Pascal University, Clermont-Ferrand
Introduction
model
Topological
navigation system
Illustrations
Discussion
38
Partial euclidian reconstruction
Unified model - Case of lines
Similar results are obtained from lines
(using polar lines)m F c F Image plane ξ
0
1
2
2
2
5 4 5 3 5 2 2 5 1 2 5 0+
+
+
+
y
+
=
A
A
x
A
A
xy
A
A
y
A
A
x
A
A
(
)
5,
,
A
A
K
h
B
i M iξ
=
Plücker coordinates
L
Unitary sphereu : unitary vector along the line L
L = (u, h, h)
T
h : unitary vector orthogonal
to the interpretation plane
the frame and the line
h : distance between the origin of
h
u
Philippe Martinet
Topological navigation using sensory memory: SLAN versus SLAM
LASMEA, Blaise Pascal University, Clermont-Ferrand
Introduction
A generic
model
Topological
navigation system
Illustrations
Discussion
39
Partial euclidian reconstruction
Unified model - Case of lines
Similar results are obtained from lines
(using polar lines)Polar lines
A
A’
Conic Φ
l
i∝ ΦA
plane of φ
A a point in the definition
φ a 2D conic curve
Polar line l
l
i: the corresponding polar line
Philippe Martinet
Topological navigation using sensory memory: SLAN versus SLAM
LASMEA, Blaise Pascal University, Clermont-Ferrand
Introduction
A generic
model
Topological
navigation system
Illustrations
Discussion
40
Partial euclidian reconstruction
Unified model - Case of lines
Similar results are obtained from lines
(using polar lines)Polar lines
m F c F ξ ur nrL
x
i
T
K
−T
ΩK
−1
x
i
= 0
Ω
i
The polar line of the optical center with respect to the conic Ω
iis given by:
Ω =
⎛
⎝
h
2 x− ξ
2¡
1
− h
2 y¢
h
xh
y¡
1
− ξ
2¢
h
xh
zh
xh
y¡
1
− ξ
2¢
h
2 y− ξ
2¡
1
− h
2 x¢
h
yh
zh
xh
zh
yh
zh
2z⎞
⎠
l
i
∝ Ω
i
O
i
O
i= [u
0v
01]
T= K[0 0 1]
Tprincipal point
Philippe MartinetTopological navigation using sensory memory: SLAN versus SLAM
LASMEA, Blaise Pascal University, Clermont-Ferrand
Introduction
A generic
model
Topological
navigation system
Illustrations
Discussion
41
Partial euclidian reconstruction
Unified model - Case of lines
Similar results are obtained from lines
(using polar lines)m F c F Image plane ξ
h
L
Unitary sphere ξVirtual image plane
Corollary:
The polar line l
icomputed from the
physical conic curve projection of L in the
omnidirectional image is the perspective
projection of L into the virtual camera
image plane.
l
i∝
K
−TΩK
−1O
i∝ K
−TΩK
−1K
⎛
⎝
0
0
1
⎞
⎠
∝
K
−T⎛
⎝
h
h
xyh
z⎞
⎠
u
Philippe MartinetTopological navigation using sensory memory: SLAN versus SLAM
LASMEA, Blaise Pascal University, Clermont-Ferrand
Introduction
A generic
model
Topological
navigation system
Illustrations
Discussion
42
Partial euclidian reconstruction
Unified model - Case of lines
Reference plane π in F
∗ mπ
∗T= [n
∗T− d
∗]
[u
∗, h
∗, h
∗] L coordinates in
F
∗ m[u, h, h] L coordinates in
Fm
Philippe Martinet
Topological navigation using sensory memory: SLAN versus SLAM
LASMEA, Blaise Pascal University, Clermont-Ferrand
Introduction
A generic
model
Topological
navigation system
Illustrations
Discussion
43
Partial euclidian reconstruction
Unified model - Case of lines
Similar results are obtained from lines
(using polar lines)X1
and
X2
are two points in the 3D space lying on the line
L
r =
h
h
∗= (1 + t
∗
d
>
R
>
n
∗
)
kn
∗×K
>l
∗ ik
kRn
∗×K
>l
ik
li
× G
−>l
∗i
= 0
with 4 couples (li, l
∗ i)
l
i∝ G
−>l
∗iR
and t
d∗=
dt∗Linear form
l
i∝ K
−TH
−TK
Tl
i∗G
= KHK
−1H
H
= R +
t d∗n
∗Tl
i∗∝ K
−Th
∗l
i∝ K
−Th
l
i∝ K
−TH
−Th
∗h
∝ H
−>h
∗ Philippe MartinetTopological navigation using sensory memory: SLAN versus SLAM
LASMEA, Blaise Pascal University, Clermont-Ferrand
Outline of the presentation
Introduction
•
LASMEA
•
Sensor based Control
•
Navigation
•
SLAN vs SLAM
Togological navigation system
•
Global system
•
Case of vision
•
Differents Modules
•
SOVIN
Discussion
•
Learning process
•
Updating process
•
Closing the loop problem
A generic model
•
A unified model for camera
•
Case of fisheye cameras
•
Partial euclidian reconstruction
Illustrations
•
Indoor applications
•
Outdoor applications
44 Philippe MartinetTopological navigation using sensory memory: SLAN versus SLAM
LASMEA, Blaise Pascal University, Clermont-Ferrand
Introduction
A generic
model
Topological
navigation system
Illustrations
Discussion
45
Global system
sensoriel Chemin Carte topologique niveau 1 Carte topologique niveau 0 Objectif courante Image Image courante LOCALISATION Image cible intermediaire Memoire sensorielle Phase hors−ligne Carte sensorielle Situation Commande courante COMMANDE CARTOGRAPHIE Sequences acquises de chemin Recherche lors de l’apprentissage Phase en−ligne Philippe MartinetTopological navigation using sensory memory: SLAN versus SLAM
LASMEA, Blaise Pascal University, Clermont-Ferrand
Introduction
A generic
model
Topological
navigation system
Illustrations
Discussion
46
Global system
sensoriel Chemin Carte topologique niveau 1 Carte topologique niveau 0 Objectif courante Image Image courante LOCALISATION Image cible intermediaire Memoire sensorielle Phase hors−ligne Carte sensorielle Situation Commande courante COMMANDE CARTOGRAPHIE Sequences acquises de chemin Recherche lors de l’apprentissage Phase en−ligneSupervised learning step: building of the sensory memory
Philippe Topological navigation using sensory memory: SLAN versus SLAM
Introduction
A generic
model
Topological
navigation system
Illustrations
Discussion
47
Global system
sensoriel Chemin Carte topologique niveau 1 Carte topologique niveau 0 Objectif courante Image Image courante LOCALISATION Image cible intermediaire Memoire sensorielle Phase hors−ligne Carte sensorielle Situation Commande courante COMMANDE CARTOGRAPHIE Sequences acquises de chemin Recherche lors de l’apprentissage Phase en−ligneInitialization: primary localization
Philippe Topological navigation using sensory memory: SLAN versus SLAM
Introduction
A generic
model
Topological
navigation system
Illustrations
Discussion
48
Global system
sensoriel Chemin Carte topologique niveau 1 Carte topologique niveau 0 Objectif courante Image Image courante LOCALISATION Image cible intermediaire Memoire sensorielle Phase hors−ligne Carte sensorielle Situation Commande courante COMMANDE CARTOGRAPHIE Sequences acquises de chemin Recherche lors de l’apprentissage Phase en−lignePhilippe Martinet
Topological navigation using sensory memory: SLAN versus SLAM
LASMEA, Blaise Pascal University, Clermont-Ferrand
Introduction
model
Topological
navigation system
Illustrations
Discussion
49
Case of vision
sensoriel Chemin Carte topologique niveau 1 Carte topologique niveau 0 Objectif couranteImage Image courante LOCALISATION Image cible intermediaire Memoire sensorielle Phase hors−ligne Carte sensorielle Situation Commande courante COMMANDE CARTOGRAPHIE Sequences acquises de chemin Recherche lors de l’apprentissage Phase en−ligne Philippe MartinetTopological navigation using sensory memory: SLAN versus SLAM
LASMEA, Blaise Pascal University, Clermont-Ferrand
Introduction
model
Topological
navigation system
Illustrations
Discussion
50
Sensory map (CS)
9Initial localization
9Sensor based control
9Vizualization of the environment
9Human interface
Topological map level 1 (CT1)
9Image network representation
9Navigability
9Commandability
9Plannification
Topological map level 0 (CT0)
9High level representation : corridor,
routes, ..
Case of vision
Sensory memory
Philippe Martinet
Topological navigation using sensory memory: SLAN versus SLAM
LASMEA, Blaise Pascal University, Clermont-Ferrand
Introduction
A generic
model
Topological
navigation system
Illustrations
Discussion
51
Different modules
Key images selection
Building the sensory memory
Learning step (first arc, first two nodes)
Philippe Martinet
Topological navigation using sensory memory: SLAN versus SLAM
LASMEA, Blaise Pascal University, Clermont-Ferrand
Introduction
A generic
model
Topological
navigation system
Illustrations
Discussion
52
Different modules
Learning step (expanding the processus…)
Philippe Martinet
Topological navigation using sensory memory: SLAN versus SLAM
LASMEA, Blaise Pascal University, Clermont-Ferrand
Introduction
A generic
model
Topological
navigation system
Illustrations
Discussion
53
Differents modules
Initial localization
How to find the closest image in the database ?
Philippe Martinet
Topological navigation using sensory memory: SLAN versus SLAM
LASMEA, Blaise Pascal University, Clermont-Ferrand
Introduction
A generic
model
Topological
navigation system
Illustrations
Discussion
54
Differents modules
Initial localization
Philippe Martinet
Topological navigation using sensory memory: SLAN versus SLAM
LASMEA, Blaise Pascal University, Clermont-Ferrand
Introduction
A generic
model
Topological
navigation system
Illustrations
Discussion
55
Differents modules
Initial localization
How to find the closest image in the database ?
Use of a regular mesh
[PERSSON04]
Cubic interpolation of the image surface
Global descriptor :
vector of the interpolated grey level
pixels at the position of the control point
Control point distribution (one example)
Philippe Martinet
Topological navigation using sensory memory: SLAN versus SLAM
LASMEA, Blaise Pascal University, Clermont-Ferrand
Introduction
A generic
model
Topological
navigation system
Illustrations
Discussion
56
Differents modules
Initial localization
How to find the closest image in the database ?
Local descriptor :
interest points detected by
Harris Stephen detector
Use of patches 11x11
Matching done by using the centered cross correlation score ZNCC
Use of the number of matched point to compute the similarity between 2 images
Philippe Martinet
Topological navigation using sensory memory: SLAN versus SLAM
LASMEA, Blaise Pascal University, Clermont-Ferrand
Introduction
A generic
model
Topological
navigation system
Illustrations
Discussion
57
Differents modules
How to find the closest image in the database ?
Evaluation of the approach [ICRA08] using three image databases regarding 3 main criteria:
Used memory size, percentage of success, computation time for localization
Several approaches were compared : GONZ [GONZALES02], PHLAC [LINAKER04],
CUB, SURF, SIFT, HARRIS, CUBHAR
Initial localization
Philippe Martinet
Topological navigation using sensory memory: SLAN versus SLAM
LASMEA, Blaise Pascal University, Clermont-Ferrand
Introduction
A generic
model
Topological
navigation system
Illustrations
Discussion
58
Differents modules
How to find the optimal path ?
Planifying the visual route
Philippe Topological navigation using sensory memory: SLAN versus SLAM
Introduction
A generic
model
Topological
navigation system
Illustrations
Discussion
59
Differents modules
How to find the optimal path ?
Planifying the visual route
Philippe Topological navigation using sensory memory: SLAN versus SLAM
Introduction
A generic
model
Topological
navigation system
Illustrations
Discussion
60
Differents modules
How to find the optimal path ?
Philippe Martinet
Topological navigation using sensory memory: SLAN versus SLAM
LASMEA, Blaise Pascal University, Clermont-Ferrand
Introduction
model
Topological
navigation system
Illustrations
Discussion
61
Differents modules
Epipolar constraint can be expressed by :
For pinhole model
Scaled euclidian reconstruction
From five couples of points
E can be estimated
Outliers are rejected by
using RANSAC
[Nister04]
X
mX
∗ mm
pm
∗ pOne line localization
Philippe Martinet
Topological navigation using sensory memory: SLAN versus SLAM
LASMEA, Blaise Pascal University, Clermont-Ferrand
Introduction
model
Topological
navigation system
Illustrations
Discussion
62
Differents modules
State estimation and control (one example)
F
Fi
= (Oi, Xi, Yi, Zi)
F
Fi+1
= (O
i+1, X
i+1, Y
i+1, Z
i+1) the frames attached to the robot when
Ii
State model of the mobile robot
Kinematic model of the mobile robot
(s, y, θ)
Chained
System
theory
[IAV04]
Philippe MartinetTopological navigation using sensory memory: SLAN versus SLAM
LASMEA, Blaise Pascal University, Clermont-Ferrand
Introduction
A generic
model
Topological
navigation system
Illustrations
Discussion
63
Differents modules
State estimation and control (one example)
Philippe Martinet
Topological navigation using sensory memory: SLAN versus SLAM
LASMEA, Blaise Pascal University, Clermont-Ferrand
Introduction
A generic
model
Topological
navigation system
Illustrations
Discussion
64
SOVIN
A global software environment for topological navigation in wide spaces
Philippe Martinet
Topological navigation using sensory memory: SLAN versus SLAM
LASMEA, Blaise Pascal University, Clermont-Ferrand
Introduction
A generic
model
Topological
navigation system
Illustrations
Discussion
65
SOVIN
A physical database organized in three levels
Physical database
Philippe Martinet
Topological navigation using sensory memory: SLAN versus SLAM
LASMEA, Blaise Pascal University, Clermont-Ferrand
Introduction
A generic
model
Topological
navigation system
Illustrations
Discussion
66
SOVIN
An efficient database interface for on line and real time applications
Philippe Martinet
Topological navigation using sensory memory: SLAN versus SLAM
LASMEA, Blaise Pascal University, Clermont-Ferrand
Introduction
A generic
model
Topological
navigation system
Illustrations
Discussion
67
SOVIN
A generic library (generic visual sensor) for visual algorithm processing
Generic image
processing library
Philippe Martinet
Topological navigation using sensory memory: SLAN versus SLAM
LASMEA, Blaise Pascal University, Clermont-Ferrand
Introduction
A generic
model
Topological
navigation system
Illustrations
Discussion
68
SOVIN
An efficient Human Machine Interface for managing applications
HMI
Philippe Martinet
Topological navigation using sensory memory: SLAN versus SLAM
LASMEA, Blaise Pascal University, Clermont-Ferrand
Outline of the presentation
Introduction
•
LASMEA
•
Sensor based Control
•
Navigation
•
SLAN vs SLAM
Togological navigation system
•
Global system
•
Case of vision
•
Differents Modules
•
SOVIN
Discussion
•
Learning process
•
Updating process
•
Closing the loop problem
A generic model
•
A unified model for camera
•
Case of fisheye cameras
•
Partial euclidian reconstruction
Illustrations
•
Indoor applications
•
Outdoor applications
69 Philippe
Martinet
Topological navigation using sensory memory: SLAN versus SLAM
LASMEA, Blaise Pascal University, Clermont-Ferrand
Introduction
A generic
model
Topological
navigation system
Illustrations
Discussion
70
Applications and projects in LASMEA
AGV using visual memory
04
05
06
07
08
09
10
11
03
AGV
Project
02
C
la
ss
ic
al
ca
m
er
a
(I
A
V
04
)
O
m
ni
di
re
ct
io
na
l
(I
R
OS
05
)
Fi
sh
ey
e
(1
,7
km
)(
IC
A
R
-C
V
08
)
WACIF OMNIBOT BODEGA CITYVIPSAFEMOVE FACT CITYHOME
µDrone
W
id
e
an
gl
e
(I
R
O
S0
9)
O
m
ni
di
re
ct
io
na
l
(I
C
R
A
07
)
C
la
ss
ic
al
ca
m
er
a
(I
C
R
A
05
)
BRI R-DISCOVER ARMENPhilippe Topological navigation using sensory memory: SLAN versus SLAM
Introduction
A generic
model
Topological
navigation system
Illustrations
Discussion
71
Indoor applications : WACIF project (2002-2005)
Autonomous Robot for Telepresence navigation, localization, learning, warning
and communication capabilities
¾
Study and development of a demonstrator like a personal
robot fully integrated in the context of « wireless home »
¾
An innovant domotic HMI dedicated for software useful
for human, family and home applications
¾
Telepresence and telesurveillance tasks
home supervision through any wireless point, abnormal
event detection in family environment (intrusion, noise, …)
Autonomous navigation strategy
Wireless communication and services
Philippe Topological navigation using sensory memory: SLAN versus SLAM
Introduction
A generic
model
Topological
navigation system
Illustrations
Discussion
72
Indoor applications : WACIF project (2002-2005)
Global navigation strategy
Room1
Room2
Topological representation
Predefined trajectory
Graph representation
Visual memory concept
Room3