ICL: ITERATIVE CLOSEST LINE/alisec/ISPRS_SS_2007/gradivo...Superposition verification Objective...
Transcript of ICL: ITERATIVE CLOSEST LINE/alisec/ISPRS_SS_2007/gradivo...Superposition verification Objective...
ICL: ITERATIVE CLOSEST LINEICL: ITERATIVE CLOSEST LINE
A NOVEL POINT CLOUD REGISTRATION A NOVEL POINT CLOUD REGISTRATION ALGORITHM BASED ON LINEAR FEATURESALGORITHM BASED ON LINEAR FEATURES
Majd ALSHAWA
1/16
Presentation outline
• Objective
• State of the art
• Proposed method
• Test and experiment
• Conclusion and future work
2/16
What is 3D matching ?
• Multiple scans for the same object• Need to put them in the same coordinate frame
Objective • match an unknown coordinate point cloud « data or scene » with a known coordinate cloud « model »
• Pairing + Rigid transformation
Particular case
Georeferenced point cloud
Topographic
scannerReal-time coordinate
acquisition
Superposition verification
Objective State of the art Proposed method Test and experiment Conclusion
3/16
State of the art
• Spheres aiding registration
• Object-based registration
“château d’eau “ details
4/16
Objective State of the art Proposed method Test and experiment Conclusion
ICP Method : Iterative Closest Point
drawback:
High number of iterations to converge (unknown approximative coordinates)
2∑ −+= iqtiRpE
• Couple the nearest points• Establish the error function :
• Minimize this function and deduce rigid transformation components• Apply this transformation to the data point cloud
Iterate the following steps :Iterate the following steps :
5/16
Objective State of the art Proposed method Test and experiment Conclusion
Proposed method: ICL : Iterative Closest Line
Data acquisition
Noise removal
ICP method
Line extraction
ICL: Form ICP
ICL : alternative Form
• ICL : an evolution of ICP
• Lines match geometric primitives
6/16
Objective State of the art Proposed method Test and experience Conclusion
Where?
Points where a remarkable normal direction change occurs
Why lines ?
• High detection possibility from many scans (invariant features)
• High registration control (two lines are sufficient)
• Reutilisation possible in other applications
7/16
Objective State of the art Proposed method Test and experiment Conclusion
Line extractionIncremental method
• Point by point modelisation
• Adding the closest point each time
• Least square adjustement to fit a line
Advantages:
• Simple and precise
• Low number of user-provided parameters
Drawback :
Execution time
8/16
Objective State of the art Proposed method Test and experiment Conclusion
Line extractionRANSAC (Random Sample Consensus)
• Trace a line through two random points
• Measure all distances
• Point number – distance criteria
• Iterate the procedure until acceptable percentage of the cloud is modeled
Advantages:
• Simple fast methodDrawback
• Probabilism
9/16
Objective State of the art Proposed method Test and experiment Conclusion
Line extractionHough Transform
Principle: line-point duality in 2D space ( coordonnates-parameters)
( )mm θρ ,( )mm θρ ,( )mm θρ , θθρ sincos YX +=10/16
Objective State of the art Proposed method Test and experiment Conclusion
COORDINATE SPACE PARAMETER SPACE
Line extraction Hough Transform
• Representing the curves on a histogram
• Regional maximums search
• Inversing the transform to detect the lines
( )mm θρ ,( )mm θρ ,( )mm θρ , θθρ sincos YX +=
advantage:
• Fast method
Drawbacks:
• Too much threshold to provide
• Incapacity in some cases
11/16
Objective State of the art Proposed method Test and experiment Conclusion
ICL Algorithm : ICP form
Iterate the following steps :
∑=
−=pN
ipixi
p
RooN
TRf1
21),(
• Line pairing
• Error function establishement :
• Rotation matrix (R) calculation
• Rotate the « data » cloud
• Shift (T) calcuation
TaRa
TaRa
+=′
+=′
22
11
1a
2a
R
′1a
′2a
T
Non
-inea
r pr
oble
mLi
nar
prob
lem
12/16
Objective State of the art Proposed method Test and experiment Conclusion
Alternative form
⎟⎟⎟
⎠
⎞
⎜⎜⎜
⎝
⎛
−−−
+⎟⎟⎟
⎠
⎞
⎜⎜⎜
⎝
⎛=
⎟⎟⎟
⎠
⎞
⎜⎜⎜
⎝
⎛+
⎟⎟⎟
⎠
⎞
⎜⎜⎜
⎝
⎛
⎟⎟⎟
⎠
⎞
⎜⎜⎜
⎝
⎛
−−−
+⎟⎟⎟
⎠
⎞
⎜⎜⎜
⎝
⎛=
⎟⎟⎟
⎠
⎞
⎜⎜⎜
⎝
⎛+
⎟⎟⎟
⎠
⎞
⎜⎜⎜
⎝
⎛
AB
AB
AB
A
A
A
T
T
T
AB
AB
AB
A
A
A
T
T
T
ZZYYXX
ZYX
ZYX
RZYX
ZZYYXX
ZYX
ZYX
RZYX
2
3
2
2
1
1
1
1
λ
λ
B’
A’
1
2
B
A
1
2
Habib et al 03
Unavoidable equation linearisation
13/16
Objective State of the art Proposed method Test and experiment Conclusion
60 mm (distance), 0.5 degree (direction)
Coupling threshold
13Paired lines
3348Extracted lines
35 mm20 mmDistance threshold
2030Number threshold
154019527Potential point number
292706601952Point number
50 mm at 60 m
30 mm at 60 mlinear resolution
I13R14Station
School of Pontonniers
R14
Two different scans
I13
Line extraction taking into account the following parameters:
RMS : 0.3 cmRMS : 0.5 m Gon
0.20.30.211
0.2-0.80.3-5
0.20.20.312
σ (cm)Shift (cm)σ (m gon)rotation (m gon)
ICL : ICP form
RMS : 0.4 cmRMS 0.3 m Gon
0.20.60.29
0.3-1.50.2-4
0.31.80.22
σ (cm)Shift (cm)σ mGonRotation( m Gon)
méthode ICL forme alternative
ICL calculation:
14/16
Objective State of the art Proposed method Test and experiment Conclusion
Conclusion
Line extraction
•Lack of an overall method encompassing all cases
•High sensitivity to thresholds
•Extracting lines in two steps: a more efficient solution
Registration •High sensibility for the last step result
•The proposed method helps to evaluate the previous topographical operations
15/16
Objective State of the art Proposed method Test and experiment Conclusion
Perspectives•Line extraction as plans contours
•Using other geometric features during the registration
•Potential point extraction study
•Supplying approximative coordinates
Positives / negatives
•Highlighting the effect of the geometric complexity
•Accelerating the registration procedure
Drawbacks• Hindrance when no protruding or recessed details exist
•Redundancy decrease according to line detection low accuracy
Advantages
16/16
Objective State of the art Proposed method Test and experiment Conclusion