topological maps based on graphs of planar regions

Topological maps based on graphs of planar regions

E. Montijano and C. Sagüés

DIIS-I3A University of Zaragoza (Spain)

IROS - 2009

Topological maps based on graphs of planar regions- IROS-2009

OUTLINE

• MOTIVATION

• SEGMENTATION OF PLANES

• CREATION OF THE TOPOLOGICAL MAPS

• EXPERIMENTAL RESULTS

• CONCLUSIONS

2

• Motivation• Segmentation• Topological Maps• Experimental Results


MOTIVATION

3


• Visual sensors provide a lot of information

• Information must be arranged• Useful for the robots• Understandable by the humans

• Hierarchical structures with images work well but• Redundant information (features repeated)• Complex structure

• In man-made environments there is abundance of planes• Easy for humans to recognize• Have been used in many robotic applications Navigation [Courbon 08] Camera calibration [Menudet 07]Visual servoing [Lopez-Nicolas 07]


4


In order to use planar regions from images

• Planar regions must be extracted

TRIPLE MATCHING APPROACH TO EXTRACT ALL THE PLANES SEEN IN A SEQUENCE OF IMAGES

• Planar regions must be sorted

EFFICIENT TOPOLOGICAL MAP CONSTRUCTION BASED ON A GRAPH OF PLANAR REGIONS

CONTRIBUTIONS


OUTLINE

5

• MOTIVATION




• CONCLUSIONS

• Segmentation• Topological Maps• Experimental Results• Conclusions


TRIPLE MATCHING (I)

6

• Initial extraction of planes in I1 and I2Homographies with DLT+RANSAC

• No information about rotation, translation or calibration

• Features close to each other are chosen as hypotheses


Starting with a sequence of images I1,…, In

• First image where the plane is seen is the reference for the plane


• Matches belonging to the plane are chosen

• ω is increased and K reduced

• The homography with respect to the reference image is also computed

• The homography is used to localize more features in the plane

7


TRIPLE MATCHING (II)

For any existing plane:


8


• New planes are extracted using the remaining matches

• Process is repeated recursively for all the images

TRIPLE MATCHING (III)

• Planes that are no longer seen are “closed”• They are not considered in the matching process anymore


9


• Multi-plane constraints make the process more robust

• Test the eigenvalues of the homologies [Sagüés 06]• Two eigenvalues equal to 1• One eigenvalue different than 1

CONSTRAINTS BETWEEN PLANES

• Test is image-based


OUTLINE

10

• MOTIVATION




• CONCLUSIONS



• The planes previously extracted are sorted in a hierarchical structure

• A graph, , A is the adjacency matrix of • Vertices represent planes in the sequence• Edges mean co-visibility between planes

•

• The matrix is symmetric• There are no self-loops

GRAPH OF PLANES

11



12


ADDITION OF PLANES

• One plane is formally added to the graph

• Each plane is represented by• Identifiers of the images where it appeared (one reference)• Features that belong to the plane• Homographies

• The process is done simultaneously with the extraction


13


FUSION OF PLANES

• Planes can leave the field of view and enter again later

• The planes would be twice in the graph

• After the creation of the graph some planes may be fused (loop closure)

• Try to compute homographies between planes not co-visible

• If an homography is computed then both planes are the same,Fusion of the planes is done by


14


OVERVIEW OF THE METHOD


OUTLINE

15

• MOTIVATION




• CONCLUSIONS



EXPERIMENTAL RESULTS

16


• Indoor sequence


17


• Outdoor sequence

EXPERIMENTAL RESULTS


OUTLINE

18

• MOTIVATION




• CONCLUSIONS



CONCLUSIONS

19

• Planar regions are useful for robots and understandable for humans

• New triple matching technique to extract the planar regions from a sequence of images has been proposed

• The planar regions have been sorted in a topological map reducing the size with respect to maps built just from images



20

¿QUESTIONS?

MORE INFO IN http://webdiis.unizar.es/~edumonti/

topological maps based on graphs of planar regions

Documents