UT Austin Villa 2013 Advances in Vision, Kinematics, and Strategy - - PowerPoint PPT Presentation

▶
ut austin villa 2013
SMART_READER_LITE
LIVE PREVIEW

UT Austin Villa 2013 Advances in Vision, Kinematics, and Strategy - - PowerPoint PPT Presentation

Nov 03, 2023 268 likes •714 views

Introduction Sequential Approach Gaussian Fitness Functions Conclusion UT Austin Villa 2013 Advances in Vision, Kinematics, and Strategy Jacob Menashe, Katie Genter, Samuel Barrett and Peter Stone

slide-1
SLIDE 1

Introduction Sequential Approach Gaussian Fitness Functions Conclusion

UT Austin Villa 2013

Advances in Vision, Kinematics, and Strategy

Jacob Menashe, Katie Genter, Samuel Barrett and Peter Stone {jmenashe,katie,sbarrett,pstone}@cs.utexas.edu The University of Texas at Austin

October 15, 2013

  • J. Menashe, K. Genter, S. Barrett, P. Stone

UT Austin Villa 2013

slide-2
SLIDE 2

Introduction Sequential Approach Gaussian Fitness Functions Conclusion

UT Austin Villa 2012 - 2013

  • J. Menashe, K. Genter, S. Barrett, P. Stone

UT Austin Villa 2013

slide-3
SLIDE 3

Introduction Sequential Approach Gaussian Fitness Functions Conclusion

Introduction

We focus on our improvements to object detection. Candidate comparison is a crucial piece of object detection Our original method compares attributes sequentially Gaussian fitness functions enable parallel evaluation

  • J. Menashe, K. Genter, S. Barrett, P. Stone

UT Austin Villa 2013

slide-4
SLIDE 4

Introduction Sequential Approach Gaussian Fitness Functions Conclusion

Introduction

We focus on our improvements to object detection. Candidate comparison is a crucial piece of object detection Our original method compares attributes sequentially Gaussian fitness functions enable parallel evaluation

  • J. Menashe, K. Genter, S. Barrett, P. Stone

UT Austin Villa 2013

slide-5
SLIDE 5

Introduction Sequential Approach Gaussian Fitness Functions Conclusion

Introduction

We focus on our improvements to object detection. Candidate comparison is a crucial piece of object detection Our original method compares attributes sequentially Gaussian fitness functions enable parallel evaluation

  • J. Menashe, K. Genter, S. Barrett, P. Stone

UT Austin Villa 2013

slide-6
SLIDE 6

Introduction Sequential Approach Gaussian Fitness Functions Conclusion

Introduction

We focus on our improvements to object detection. Candidate comparison is a crucial piece of object detection Our original method compares attributes sequentially Gaussian fitness functions enable parallel evaluation

  • J. Menashe, K. Genter, S. Barrett, P. Stone

UT Austin Villa 2013

slide-7
SLIDE 7

Introduction Sequential Approach Gaussian Fitness Functions Conclusion Sequential Sanity Checks Problems

Sequential Sanity Checks for Object Detection

  • J. Menashe, K. Genter, S. Barrett, P. Stone

UT Austin Villa 2013 1

Retrieve candidates with blob detection

Image from https://maserati.mi.fu-berlin.de/fub-kit/?tag=robocup-2013

slide-8
SLIDE 8

Introduction Sequential Approach Gaussian Fitness Functions Conclusion Sequential Sanity Checks Problems

Sequential Sanity Checks for Object Detection

  • J. Menashe, K. Genter, S. Barrett, P. Stone

UT Austin Villa 2013 1

Retrieve candidates with blob detection

2

Sanity check each candidate

Image from https://maserati.mi.fu-berlin.de/fub-kit/?tag=robocup-2013

slide-9
SLIDE 9

Introduction Sequential Approach Gaussian Fitness Functions Conclusion Sequential Sanity Checks Problems

Sequential Sanity Checks for Object Detection

  • J. Menashe, K. Genter, S. Barrett, P. Stone

UT Austin Villa 2013 1

Retrieve candidates with blob detection

2

Sanity check each candidate

3

Accept the first candidate to pass all tests

Image from https://maserati.mi.fu-berlin.de/fub-kit/?tag=robocup-2013

slide-10
SLIDE 10

Introduction Sequential Approach Gaussian Fitness Functions Conclusion Sequential Sanity Checks Problems

Problems with the Sequential Approach

For simple detection problems, the sequential approach works well

  • J. Menashe, K. Genter, S. Barrett, P. Stone

UT Austin Villa 2013

slide-11
SLIDE 11

Introduction Sequential Approach Gaussian Fitness Functions Conclusion Sequential Sanity Checks Problems

Problems with the Sequential Approach

For simple detection problems, the sequential approach works well Simple to code, test and modify

  • J. Menashe, K. Genter, S. Barrett, P. Stone

UT Austin Villa 2013

slide-12
SLIDE 12

Introduction Sequential Approach Gaussian Fitness Functions Conclusion Sequential Sanity Checks Problems

Problems with the Sequential Approach

For simple detection problems, the sequential approach works well Simple to code, test and modify More advanced scenarios can be problematic

  • J. Menashe, K. Genter, S. Barrett, P. Stone

UT Austin Villa 2013

slide-13
SLIDE 13

Introduction Sequential Approach Gaussian Fitness Functions Conclusion Sequential Sanity Checks Problems

Distinguishing Between Candidates

  • J. Menashe, K. Genter, S. Barrett, P. Stone

UT Austin Villa 2013 Image from http://www.intechopen.com/books/robot-soccer/humanoid-soccer-player-design

slide-14
SLIDE 14

Introduction Sequential Approach Gaussian Fitness Functions Conclusion Sequential Sanity Checks Problems

Determining Detection Quality

  • J. Menashe, K. Genter, S. Barrett, P. Stone

UT Austin Villa 2013

slide-15
SLIDE 15

Introduction Sequential Approach Gaussian Fitness Functions Conclusion Sequential Sanity Checks Problems

Parallel Parameter Evaluation

  • Blue bars are readings, red lines are thresholds.
  • J. Menashe, K. Genter, S. Barrett, P. Stone

UT Austin Villa 2013

slide-16
SLIDE 16

Introduction Sequential Approach Gaussian Fitness Functions Conclusion Solution Method Example

Solution: Multivariate Gaussian Fitness Functions

  • J. Menashe, K. Genter, S. Barrett, P. Stone

UT Austin Villa 2013

Sanity checks are performed simultaneously

Image from http://en.wikipedia.org/wiki/Gaussian_function

slide-17
SLIDE 17

Introduction Sequential Approach Gaussian Fitness Functions Conclusion Solution Method Example

Solution: Multivariate Gaussian Fitness Functions

  • J. Menashe, K. Genter, S. Barrett, P. Stone

UT Austin Villa 2013

Sanity checks are performed simultaneously Output is a float in [0,1]

Image from http://en.wikipedia.org/wiki/Gaussian_function

slide-18
SLIDE 18

Introduction Sequential Approach Gaussian Fitness Functions Conclusion Solution Method Example

Solution: Multivariate Gaussian Fitness Functions

  • J. Menashe, K. Genter, S. Barrett, P. Stone

UT Austin Villa 2013

Sanity checks are performed simultaneously Output is a float in [0,1] Fitness scores are directly comparable

Image from http://en.wikipedia.org/wiki/Gaussian_function

slide-19
SLIDE 19

Introduction Sequential Approach Gaussian Fitness Functions Conclusion Solution Method Example

Method Overview

1

Select measurements

  • J. Menashe, K. Genter, S. Barrett, P. Stone

UT Austin Villa 2013

slide-20
SLIDE 20

Introduction Sequential Approach Gaussian Fitness Functions Conclusion Solution Method Example

Method Overview

1

Select measurements

2

Determine the mean µ and covariance matrix Σ

  • J. Menashe, K. Genter, S. Barrett, P. Stone

UT Austin Villa 2013

slide-21
SLIDE 21

Introduction Sequential Approach Gaussian Fitness Functions Conclusion Solution Method Example

Method Overview

1

Select measurements

2

Determine the mean µ and covariance matrix Σ

3

Use measurements to compute a feature vector v

  • J. Menashe, K. Genter, S. Barrett, P. Stone

UT Austin Villa 2013

slide-22
SLIDE 22

Introduction Sequential Approach Gaussian Fitness Functions Conclusion Solution Method Example

Method Overview

1

Select measurements

2

Determine the mean µ and covariance matrix Σ

3

Use measurements to compute a feature vector v

4

Compute fitness f using µ, Σ, and the multivariate Gaussian PDF G

  • J. Menashe, K. Genter, S. Barrett, P. Stone

UT Austin Villa 2013

slide-23
SLIDE 23

Introduction Sequential Approach Gaussian Fitness Functions Conclusion Solution Method Example

Method Overview

1

Select measurements

2

Determine the mean µ and covariance matrix Σ

3

Use measurements to compute a feature vector v

4

Compute fitness f using µ, Σ, and the multivariate Gaussian PDF G f = G(v; µ, Σ)/G(µ; µ, Σ)

  • J. Menashe, K. Genter, S. Barrett, P. Stone

UT Austin Villa 2013

slide-24
SLIDE 24

Introduction Sequential Approach Gaussian Fitness Functions Conclusion Solution Method Example

Measurement Selection

  • J. Menashe, K. Genter, S. Barrett, P. Stone

UT Austin Villa 2013

Velocity Orange % Green/White % Circle Deviation Field Distance Perceived Height Distance Discrepancy

slide-25
SLIDE 25

Introduction Sequential Approach Gaussian Fitness Functions Conclusion Solution Method Example

Measurement Selection

  • J. Menashe, K. Genter, S. Barrett, P. Stone

UT Austin Villa 2013

v = (0, . . .)⊤

Velocity Orange % Green/White % Circle Deviation Field Distance Perceived Height Distance Discrepancy

slide-26
SLIDE 26

Introduction Sequential Approach Gaussian Fitness Functions Conclusion Solution Method Example

Measurement Selection

  • J. Menashe, K. Genter, S. Barrett, P. Stone

UT Austin Villa 2013

v = (0, .95, . . .)⊤

Velocity Orange % Green/White % Circle Deviation Field Distance Perceived Height Distance Discrepancy

slide-27
SLIDE 27

Introduction Sequential Approach Gaussian Fitness Functions Conclusion Solution Method Example

Measurement Selection

  • J. Menashe, K. Genter, S. Barrett, P. Stone

UT Austin Villa 2013

v = (0, .95, .93, . . .)⊤

Velocity Orange % Green/White % Circle Deviation Field Distance Perceived Height Distance Discrepancy

slide-28
SLIDE 28

Introduction Sequential Approach Gaussian Fitness Functions Conclusion Solution Method Example

Measurement Selection

  • J. Menashe, K. Genter, S. Barrett, P. Stone

UT Austin Villa 2013

v = (0, .95, .93, .18, . . .)⊤

Velocity Orange % Green/White % Circle Deviation Field Distance Perceived Height Distance Discrepancy

slide-29
SLIDE 29

Introduction Sequential Approach Gaussian Fitness Functions Conclusion Solution Method Example

Measurement Selection

  • J. Menashe, K. Genter, S. Barrett, P. Stone

UT Austin Villa 2013

v = (0, .95, .93, .18, 50, . . .)⊤

Velocity Orange % Green/White % Circle Deviation Field Distance Perceived Height Distance Discrepancy

slide-30
SLIDE 30

Introduction Sequential Approach Gaussian Fitness Functions Conclusion Solution Method Example

Measurement Selection

  • J. Menashe, K. Genter, S. Barrett, P. Stone

UT Austin Villa 2013

v = (0, .95, .93, .18, 50, 65, . . .)⊤

Velocity Orange % Green/White % Circle Deviation Field Distance Perceived Height Distance Discrepancy

slide-31
SLIDE 31

Introduction Sequential Approach Gaussian Fitness Functions Conclusion Solution Method Example

Measurement Selection

  • J. Menashe, K. Genter, S. Barrett, P. Stone

UT Austin Villa 2013

v = (0, .95, .93, .18, 50, 65, . . .)⊤

Velocity Orange % Green/White % Circle Deviation Field Distance Perceived Height Distance Discrepancy

slide-32
SLIDE 32

Introduction Sequential Approach Gaussian Fitness Functions Conclusion Solution Method Example

Measurement Selection

  • J. Menashe, K. Genter, S. Barrett, P. Stone

UT Austin Villa 2013

v = (0, .95, .93, .18, 50, 65, . . .)⊤

Velocity Orange % Green/White % Circle Deviation Field Distance Perceived Height Distance Discrepancy

slide-33
SLIDE 33

Introduction Sequential Approach Gaussian Fitness Functions Conclusion Solution Method Example

Measurement Selection

  • J. Menashe, K. Genter, S. Barrett, P. Stone

UT Austin Villa 2013

v = (0, .95, .93, .18, 50, 65, .2)⊤

Velocity Orange % Green/White % Circle Deviation Field Distance Perceived Height Distance Discrepancy

slide-34
SLIDE 34

Introduction Sequential Approach Gaussian Fitness Functions Conclusion Solution Method Example

Gaussian Parameters

Measurement µ σ Velocity 0.0 max(100, d/5) Orange Percentage 1.0 0.5 Green/White Percentage 1.0 0.4 Circle Deviation 0.0 0.3 Field Distance 0.0 max(100, d/10) Perceived Height 0.0 150.0 Distance Discrepancy 0.0 0.4 d is the last known ball distance Σ computed as the diagonal matrix with entries σ2 from each measurement

  • J. Menashe, K. Genter, S. Barrett, P. Stone

UT Austin Villa 2013

slide-35
SLIDE 35

Introduction Sequential Approach Gaussian Fitness Functions Conclusion Solution Method Example

Computing Fitness

f = G(v, µ, Σ) G(µ, µ, Σ = G                     .95 .93 .18 50 65 .2           ,           1 1           ,           5002 .52 .42 .32 2502 1502 .42                     G(µ, µ, Σ) = 0.65

  • J. Menashe, K. Genter, S. Barrett, P. Stone

UT Austin Villa 2013

slide-36
SLIDE 36

Introduction Sequential Approach Gaussian Fitness Functions Conclusion Solution Method Example

Experiment: Baseline

  • J. Menashe, K. Genter, S. Barrett, P. Stone

UT Austin Villa 2013

Fitness scores and sequential selections from 500 frames at static

  • positions. The object of

interest (left) and decoy (right) are nearly identical.

Left µ, σ .92, .04 Right µ, σ .89, .02 Gaussian P(success) .7291 Sequential P(success) .68

slide-37
SLIDE 37

Introduction Sequential Approach Gaussian Fitness Functions Conclusion Solution Method Example

Experiment: Velocity

  • J. Menashe, K. Genter, S. Barrett, P. Stone

UT Austin Villa 2013

Fitness scores and sequential selections from 500 frames at static

  • positions. The object of

interest (left) and decoy (right) differ in their computed velocities.

Left µ, σ .88, .12 Right µ, σ .15, .20 Gaussian P(success) .9992 Sequential P(success) 0.00

slide-38
SLIDE 38

Introduction Sequential Approach Gaussian Fitness Functions Conclusion Solution Method Example

Experiment: Height

  • J. Menashe, K. Genter, S. Barrett, P. Stone

UT Austin Villa 2013

Fitness scores and sequential selections from 500 frames at static

  • positions. The object of

interest (left) and decoy (right) differ only in height.

Left µ, σ .90, .02 Right µ, σ .78, .04 Gaussian P(success) .9965 Sequential P(success) .87

slide-39
SLIDE 39

Introduction Sequential Approach Gaussian Fitness Functions Conclusion Solution Method Example

Experiment: Size

  • J. Menashe, K. Genter, S. Barrett, P. Stone

UT Austin Villa 2013

Fitness scores and sequential selections from 500 frames at static

  • positions. The object of

interest (left) and decoy (right) differ only in size.

Left µ, σ .89, .04 Right µ, σ .39, .01 Gaussian P(success) > .9999 Sequential P(success) 1.00

slide-40
SLIDE 40

Introduction Sequential Approach Gaussian Fitness Functions Conclusion

Related Work

Samuel Barrett, Katie Genter, Todd Hester, Piyush Khandelwal, Michael Quinlan, Peter Stone, and Mohan

  • Sridharan. Austin Villa 2011: Sharing is caring: Better awareness through information sharing. Technical Report

UT-AI-TR-12-01, The University of Texas at Austin, Department of Computer Sciences, AI Laboratory, January 2012. Morgan Quigley, Ken Conley, Brian Gerkey, Josh Faust, Tully B. Foote, Jeremy Leibs, Rob Wheeler, and Andrew Y.

  • Ng. ROS: an open-source robot operating system. In ICRA Workshop on Open Source Software, 2009.

Thomas Röfer, Tim Laue, Judith Müller, Alexander Fabisch, Fynn Feldpausch, Katharina Gillmann, Colin Graf, Thijs Jeffry de Haas, Alexander Härtl, Arne Humann, Daniel Honsel, Philipp Kastner, Tobias Kastner, Carsten Könemann, Benjamin Markowsky, Ole Jan Lars Riemann, and Felix Wenk. B-Human team report and code release, 2011. http://www.b-human.de/downloads/bhuman11_coderelease.pdf. Camiel Verschoor, Auke Wiggers, Duncan ten Velthuis, Anna Keune, Michael Cabot, Sander Nugteren, Erik van Egmond, Hessel van der Molen, Robert Iepsma, Maurits van Bellen, Merel de Groot, Eszter Fodor, Richard Rozeboom, and Arnoud Visser. Dutch nao team - technical report, 2011.

  • J. Menashe, K. Genter, S. Barrett, P. Stone

UT Austin Villa 2013

slide-41
SLIDE 41

Introduction Sequential Approach Gaussian Fitness Functions Conclusion

Conclusion

  • J. Menashe, K. Genter, S. Barrett, P. Stone

UT Austin Villa 2013

Improved on the sequential approach with Gaussian fitness functions

slide-42
SLIDE 42

Introduction Sequential Approach Gaussian Fitness Functions Conclusion

Conclusion

  • J. Menashe, K. Genter, S. Barrett, P. Stone

UT Austin Villa 2013

Improved on the sequential approach with Gaussian fitness functions Described the implementation details for the case of ball detection

slide-43
SLIDE 43

Introduction Sequential Approach Gaussian Fitness Functions Conclusion

Conclusion

  • J. Menashe, K. Genter, S. Barrett, P. Stone

UT Austin Villa 2013

Improved on the sequential approach with Gaussian fitness functions Described the implementation details for the case of ball detection This method can be applied for a variety of other field objects

slide-44
SLIDE 44

Introduction Sequential Approach Gaussian Fitness Functions Conclusion

Conclusion

  • J. Menashe, K. Genter, S. Barrett, P. Stone

UT Austin Villa 2013

Improved on the sequential approach with Gaussian fitness functions Described the implementation details for the case of ball detection This method can be applied for a variety of other field objects Improvements to kinematics and strategy included in our work.