CmpE-537 Computer Vision

1
CmpE-537 Computer Vision

• Term Project
• Color and Illumination Independent Landmark
  Detection for Robot Soccer Domain
• By
• Tekin Meriçli
• Artificial Intelligence Laboratory
• Department of Computer Engineering
• Bogaziçi University
• 27/12/2007

2
Outline

• Introduction
• Related Work
• Proposed Approach
• Experimental Setup
• Results
• Conclusion
• References

3
Introduction

• Three fundamental questions of mobile robotics
• Where am I?,
• Where am I going?,
• How can I get there?
• The aim of this project is to answer the ?rst
  question for robot soccer domain
• Speci?cally RoboCup Standard Platform League
  (former 4-Legged League)?
• Robots with vision sensors (i.e. cameras) are used

4
Introduction
5
Introduction

• All important objects on the ?eld, that is the
  ball, the beacons, and the goals, are color-coded
• This makes vision, and hence localization modules
  highly dependent on illimunation
• The robots may not be able to detect the beacons
  at all, or calculate their distances and
  orientations to the beacons wrong if there is
  even a small change in the illumination level
• Main motivation is to make the vision /
  localization processes color and illumination
  independent in the Standard Platform League domain

6
Related Work

• Color / illumination dependent approach
• Color segmentation / pixel classi?cation on the
  image
• Connected component analysis to build regions
• Sanity checks to remove noise and illogical
  perceptions
• aspect ratio, minimum area, etc.
• Most of the RoboCup teams use this approach 14

7
Related Work

• Feature detection / recognition based approach
• Used for simultaneous localization and mapping
  (SLAM) purposes
• Scale-invariant feature transform (or SIFT) can
  be used in algorithms for tasks like matchin
  different views of an object or scene (e.g. for
  stereo vision) and object recognition 7
• SURF, which stands for Speeded-Up Robust
  Features, approximates SIFT

8
Proposed Approach

• Image labeling process that has been used in
  color segmentation-based approach is replaced
  with region labeling in which the landmarks and
  their immediate surrounding are covered
• The robot is placed at a location where it can
  see the landmark, and then a region is selected
  around the landmark to specify the region in
  which the robot should ?nd the SURF features and
  associate them with that particular landmark

9
Proposed Approach
10
Proposed Approach

• This process is repeated for all landmarks on the
  soccer ?eld from different angles and distances
• Supervised learning is used to learn the
  associations between the feature descriptors and
  the landmarks
• The distance values for landmarks are calculated
  using the inter-feature distances

11
Experimental Setup

• A real Aibo ERS-7 robot is placed on the field
  facing a particular landmark with different
  angles and distances to take pictures
• An offline visualizer tool is implemented to show
  the SURF points on the image and run tests on
  various images

12
Experimental Setup
13
Experimental Setup

• SURF points are shown as little circles
• Details of the descriptors are listed on the text
  area
• Similar feature points are observed on different
  images even though the distance and angle values
  are different
• Similarity is defined as the distance between
  feature points in 64 dimensional feature space

14
Experimental Setup

• First step is to process the training images and
  define the landmark regions by clicking on the
  image
• The next step is to test images to check whether
  the landmark in the image is recognized and
  whether the distance and angle estimates are
  correct

15
Results

• SURF computation took an average of 56ms on
  354x290 images
• Aibo robots capture 208x160 images, but have a
  slower processor hence, SURF computation takes
  59ms on average, which is approximately 17fps
• Landmark recognition performance was better than
  distance estimates
• Due to the cylindrical shape of landmarks, some
  feature points may be closer to or farther from
  each other depending on the angle, or may totally
  be hidden
• Doing the computations on groups of feature
  points rather than individuals may improve the
  performance

16
Conclusion

• A feature-based landmark detection approach is
  explored
• Runs with reasonable fps rate
• Main contribution is that this approach provides
  color (and illumination to some extent)
  independence in vision and localization processes
  in robot soccer domain
• It has not been tried by any of the RoboCup teams
  so far
• Trying different SURF parameters and running
  experiments on physical robots are left as future
  work

17
References

• 1 H. L. Akin et.al. Cerberus 2006 Team
  Report. 2006.
• 2 Kaplan, K., B. Celik, T. Mericli, C. Mericli,
  and H. L. Akin. Practical Extensions to
  Vision-Based Monte Carlo Localization Methods for
  Robot Soccer Domain, In RoboCup International
  Symposium 2005, Osaka, July 18-19, 2005.
• 3 Peter Stone, Peggy Fidelman, Nate Kohl,
  Gregory Kuhlmann, Tekin Mericli, Mohan Sridharan,
  and Shao-en Yu. The UT Austin Villa 2006 RoboCup
  Four-Legged Team. Technical Report
  UT-AI-TR-06-337, The University of Texas at
  Austin, Department of Computer Sciences, AI
  Laboratory, 2006.
• 4 M. J. Quinlan et.al. The 2006 NUbots Team
  Report, 2007.
• 5 Thomas Roefer et.al. GermanTeam2006, 2006.
• 6 Herbert Bay, Tinne Tuytelaars, Luc J. Van
  Gool. SURF Speeded Up Robust Features, In
  ECCV06, pp.404-417, 2006.
• 7 Lowe, D. G., Distinctive Image Features from
  Scale-Invariant Keypoints, In International
  Journal of Computer Vision, 60, 2, pp. 91-110,
  2004.

18
References

• 8 M. Ballesta, A. Gil, O. Martnez Mozos, and O.
  Reinoso. Local descriptors for visual slam. In
  Proc. of the Workshop on Robotics and
  Mathematics, Coimbra, Portugal, 2007.
• 9 Barfoot, T D, Online Visual Motion
  Estimation using FastSLAM with SIFT Features. In
  Proc. of the Int. Conf. on Robotics and
  Intelligent Systems (IROS), Edmonton, Alberta,
  August 2-6, 2005.
• 10 Pantelis Elinas and James J. Little. Stereo
  vision SLAM Near real-time learning of 3D
  point-landmark and 2D occupancy-grid maps using
  particle lters. In IROS07, 2007.
• 11 J. Little, S. Se, and D.G. Lowe.
  Vision-based mobile robot localization and
  mapping using scale-invariant features. In IEEE
  Int. Conf. on Robotics Automation, 2001.
• 12 Martnez Mozos, O. and Gil, A. and Ballesta,
  M. and Reinoso, O. Interest Point Detectors for
  Visual SLAM. In Lecture Notes in Arti?cial
  Intelligence, vol4788, 2007.

19
?