+ All Categories
Home > Documents > Robust Extraction of Vertices in Range Images by Constraining the Hough Transform

Robust Extraction of Vertices in Range Images by Constraining the Hough Transform

Date post: 30-Jan-2016
Category:
Upload: indra
View: 33 times
Download: 0 times
Share this document with a friend
Description:
Robust Extraction of Vertices in Range Images by Constraining the Hough Transform. Dimitrios Katsoulas Institute for Pattern Recognition and Image Processing University of Freiburg. Introduction. Depalletizing : automatic unloading of piled objects via a robot . - PowerPoint PPT Presentation
32
Robust Extraction of Robust Extraction of Vertices in Range Images by Vertices in Range Images by Constraining the Hough Constraining the Hough Transform Transform Dimitrios Katsoulas Institute for Pattern Recognition and Image Processing University of Freiburg
Transcript
Page 1: Robust Extraction of Vertices in Range Images by Constraining the Hough Transform

Robust Extraction of Vertices in Robust Extraction of Vertices in Range Images by Constraining the Range Images by Constraining the

Hough TransformHough Transform

Dimitrios KatsoulasInstitute for Pattern Recognition and Image Processing

University of Freiburg

Page 2: Robust Extraction of Vertices in Range Images by Constraining the Hough Transform

IntroductionIntroduction

Depalletizing: automatic unloading of piled objects via a robot.

How important is a solution to the problem? Applications: Post, distribution centres,

airports. We deal with objects most frequently

encountered: Boxes, box-like objects (e.g. sacks full of material)

Page 3: Robust Extraction of Vertices in Range Images by Constraining the Hough Transform

Object configurationsObject configurations

Page 4: Robust Extraction of Vertices in Range Images by Constraining the Hough Transform

Construction of an automatic Construction of an automatic systemsystem

Sensor: Range sensor for data acquisition, since it is of advantage to have depth information.

Modelling: geometric parametric models: superquadrics.

Recovery: Based on hypothesis generation and refinement.

[ChenKak 89]: 3D Vertices provide the strongest constraints for generating hypotheses about object pose in space Vertex detection of extreme importance for hypothesis generation!!

Page 5: Robust Extraction of Vertices in Range Images by Constraining the Hough Transform

Vertex detection in ImagesVertex detection in Images Many approaches for detecting corners in intensity

images, e.g. [Harris.et.al 88], [Smith.et.al 98] (SUSAN), [Deriche.et.al 92], [Zhang.et.al 94]

In range images: The majority employs region information e.g. [ChenKak 89], [Baerveldt 93].

Problem: More than one surface needs to be exposed to the range sensor.

Our approach: Utilizes boundary information, based on Hough Transform:

1. Image acquisition2. Linear boundary detection in 3D3. Boundary grouping

Page 6: Robust Extraction of Vertices in Range Images by Constraining the Hough Transform

ROADMAPROADMAP

Data acquisition Detection of linear object boundaries in 3D

– Edge detection – Parameter recovery with the Hough Transform– Line Detection in 3D– Model Selection

Boundary grouping Experiments Discussion

Page 7: Robust Extraction of Vertices in Range Images by Constraining the Hough Transform

Data acquisitionData acquisition

Page 8: Robust Extraction of Vertices in Range Images by Constraining the Hough Transform

ROADMAPROADMAP

Data acquisition Detection of linear object boundaries in 3D

– Edge detection – Parameter recovery with the Hough Transform– Line Detection in 3D– Model Selection

Boundary grouping Experiments Discussion

Page 9: Robust Extraction of Vertices in Range Images by Constraining the Hough Transform

Edge map generationEdge map generation

We use the detector of [Jiang.et.al 00] which approximates the scan lines with linear and parabolic segments.

Advantage with regard to local edge detectors: Accuracy, ability to detect crease edges.

Problem: Error in localizing edge points, models do not perfectly express the objects.

Page 10: Robust Extraction of Vertices in Range Images by Constraining the Hough Transform

Output of edge detectionOutput of edge detection

Edges_3D_out.pvboxes_out.pv

How can we detect linear boundaries from an edge map?

Page 11: Robust Extraction of Vertices in Range Images by Constraining the Hough Transform

Parameter recovery with the Parameter recovery with the Standard Hough TransformStandard Hough Transform

Line Equation: x cos( ) ysin( )

y

x

Page 12: Robust Extraction of Vertices in Range Images by Constraining the Hough Transform

Problems of the SHTProblems of the SHT Localization error not addressed! Computationally inefficient and memory consuming:

– If N the dof of the model sought– p the number of parameters constrained by each point in

the image space– Then mapping of each image point requires incrementing

all bins comprising a N-p - dimensional manifold of the accumulator.

Test case 3D lines: N = 4, p = 2 4D accumulator, update of a 2D manifold per mapping.

Page 13: Robust Extraction of Vertices in Range Images by Constraining the Hough Transform

Improving the performance of Hough Improving the performance of Hough transform by problem decompositiontransform by problem decomposition

Idea: Decompose the transform into sub-problems:– Select set of points with cardinality d (d<N)– Consider r random subsets of the remaining points

with cardinality v, so that d + v = N– Map the union to the parameter space.– Perform t trials.– Find peaks.

Adjustment of d,r,t results to a variety of approaches: Most representative [Leavers 92] , [Olson 00], [Olson 01].

TRIAL

Page 14: Robust Extraction of Vertices in Range Images by Constraining the Hough Transform

RRecognition ecognition uusing sing DDecomposition ecomposition and and RRandomization [Olson 01]andomization [Olson 01]

Parameter selection:– d = N -1 – r = n -1, where n the number of edge points in

image

– If the user defined probability of failure in finding a model and m the minimum number of expected to lie on the model then:

γ

N 1

1log( )

tmn

Page 15: Robust Extraction of Vertices in Range Images by Constraining the Hough Transform

Constraining the Hough Constraining the Hough TransformTransform

Line equation: x cos( ) ysin( )

y

x

Page 16: Robust Extraction of Vertices in Range Images by Constraining the Hough Transform

Advantages of RUDRAdvantages of RUDR

Complexity: O(tr) t: constant O(n)By selecting d = N - 1:

– The transform is constrained to lie on a 1D curve, the Hough Curve

– 1D data structures are used for accumulation Memory requirements O(n)

– Accurate error propagation in the parameter space.

Page 17: Robust Extraction of Vertices in Range Images by Constraining the Hough Transform

Error propagationError propagation

Page 18: Robust Extraction of Vertices in Range Images by Constraining the Hough Transform

Line detection in 3DLine detection in 3D

3D case: points constrain two line parameters.

Thereby, we cannot constrain the transform to lie on a curve.

Solution: Break down the problem in two 2D sub-problems:– Determine 2 line parameters on the image (XZ)

plane.– Determine the remaining on the plane defined by

the detected line and the Y axis.

Page 19: Robust Extraction of Vertices in Range Images by Constraining the Hough Transform

Decomposition in two 2D sub-Decomposition in two 2D sub-problemsproblems

Page 20: Robust Extraction of Vertices in Range Images by Constraining the Hough Transform

Flow diagram of a TrialFlow diagram of a Trial

Page 21: Robust Extraction of Vertices in Range Images by Constraining the Hough Transform

The problem of redundancyThe problem of redundancy Our algorithm does not guarantee that only one

point of the same boundary is used as distinguished point.

Point removal is problematic. Instead of retaining a locally sufficient model we:

– Wait until all trials take place– Retain models satisfying global optimality criteria

Criteria: Favor models which describe bigger number of image points with higher accuracy.

Selection via MDL

Page 22: Robust Extraction of Vertices in Range Images by Constraining the Hough Transform

Selecting the optimal linesSelecting the optimal lines

If M models recovered, we consider the matrix: and the binary vector

Diagonal elements:

Off- diagonals:

Benefit function:

Model recovery:

MxMQii 1 i 2 iq K M K V

1 i j

ij

K M Mq

2

TF(m) m Qm% % %

mmax F(m)

% %

m%

Page 23: Robust Extraction of Vertices in Range Images by Constraining the Hough Transform

Selecting optimal lines (2)Selecting optimal lines (2)

Maximization though simulated annealing or neural networks computationally inefficient.

We adopt the greedy algorithm of proposed in [jaklic et al 00]

2O(M )

Page 24: Robust Extraction of Vertices in Range Images by Constraining the Hough Transform

Effect of Model SelectionEffect of Model Selection

beforeSelection2_out.pv afterSelection2_out.pv

Page 25: Robust Extraction of Vertices in Range Images by Constraining the Hough Transform

The Detection-Selection The Detection-Selection processprocess

Complexity:

2O(tn) O(M ) O(n)

since

t O(1) n M ?

Page 26: Robust Extraction of Vertices in Range Images by Constraining the Hough Transform

ROADMAPROADMAP

Data acquisition Detection of linear object boundaries in 3D

– Edge detection– Parameter recovery with the Hough Transform– Line Detection in 3D– Model Selection

Boundary grouping Experiments Discussion

Page 27: Robust Extraction of Vertices in Range Images by Constraining the Hough Transform

Boundary groupingBoundary grouping 3D vertex: Aggregate consisting of:

– Two orthogonal 3D line segments.– Vertex point.

Ideally: If X,Y linear segments the dot product of their direction vectors should be zero:

Threshold: difficult to set, depends on the application and on the uncertainty in calculation of line parameters.

Can we avoid multiple thresholds? Yes by introducing statistical tests.

Tr x y 0 %%

x%

y%

Page 28: Robust Extraction of Vertices in Range Images by Constraining the Hough Transform

Statistical testStatistical testinging the the geometric relationgeometric relation

We adopt the framework of [Foerstner 00] for its compactness and straightforwardness.

Dot product: Bilinear function of two stochastic vectors:

The variance of their dot product is given by: 2 T Tr yy xxx x y y

%%%%% % % %

x xxx ~ N( , ) % %%%y yyy ~ N( , ) % %%%

Page 29: Robust Extraction of Vertices in Range Images by Constraining the Hough Transform

Statistical testStatistical testinging the the geometric relationgeometric relation (2) (2)

The optimal test statistic for the hypothesis: is given by:

A significance value is selected. If the hypothesis that the relation holds can be rejected. Overall algorithm for boundary grouping:

– Consider all pairs of lines.– The pairs passing the test along their intersection

points are the detected vertices.

r

rt N(0,1)

:

0H : r 0

1t N (0,1)

Page 30: Robust Extraction of Vertices in Range Images by Constraining the Hough Transform

Experiments (1)Experiments (1)

Edges_3D_out.pv Lines3D_out.pv vertices3D_out.pv

Page 31: Robust Extraction of Vertices in Range Images by Constraining the Hough Transform

Experiments (2)Experiments (2)

Edges3D_out.pv Lines3D_out.pv vertices3D_out.pv

Page 32: Robust Extraction of Vertices in Range Images by Constraining the Hough Transform

DiscussionDiscussion System advantages:

– Robustness: Error propagation, statistical test for grouping

– Accuracy: Model selection.– Computational efficiency: Linear complexity to the

number of edge points.– Low memory consumption: 1D accumulators used.– Versatility: Deals with both piled and neatly placed

configurations.– Simplicity.

System drawbacks: What if the objects are placed very close one after another?


Recommended