Machine Vision Seminar - PowerPoint PPT Presentation

View by Category
About This Presentation
Title:

Machine Vision Seminar

Description:

As part of Machine Vision seminar, I had to present a paper. This is the presentation I used. – PowerPoint PPT presentation

Number of Views:2325
Updated: 5 March 2011
Slides: 38
Provided by: faiss

less

Write a Comment
User Comments (0)
Transcript and Presenter's Notes

Title: Machine Vision Seminar


1
A Real-Time Algorithm for Mobile Robot Mapping
With Applications to Multi-Robot and 3D Mapping
  • Sebastian Thrun Wolfram Burgard
    Dieter Fox

Presented by Mohammad Faisal Smart
Systems Spring 2011 Machine Vision Seminar

Presenter Notes
10000 papers per year 100 per presentation..
Many not important this one EXTREMELY

Won awards..

Relavant now as it was before

Introduction of author. Fast Slam.. Numerous
awards.. Google singularity uni

2
Structure
  • Motivation
  • Existing Work
  • Approach
  • Details
  • Results
  • Conclusion

3
Motivation
  • Importance of maps
  • Importance of localization
  • Slam


Presenter Notes
Maps are used to know where to go and where not
to, Help us know where we are, Useful in
construction, Danger accessment





Localization more important. Know about yourself
before you judge others



Tunnel coming out in middle of rome find out
where you are..

Approach?



Simultaneous Localisation and Mapping



4
SLAM
  • What makes Slam difficult
  • Map available
  • 1996
  • map matching
  • Localization available
  • 1984
  • Using Sonar


Presenter Notes
A whole book match the scan againnst existing
map

Local localisation global localisation



Sonar.. Images from different positions
positions known to each other occupancy grid.



5
Kalman Filter
  • Iterative
  • Failure
  • Cyclic environment
  • Cumulative Error
  • No Backwards Correction


Presenter Notes
Kalman filter.. Combining sensor inputs refining
the error function..

Predict where you will be Check if you are
there AND ITERATE

6
EM - SLAM
  • Expectation Maximization
  • Simultaneously all past scans
  • Iterative refinement
  • Cycles actually make things more accurate
  • Disadvantages
  • Batch processes
  • Offline


Presenter Notes
Maximum Likelihood.

constrained, probabilistic maximumlikelihood
estimation problem

7
Simplicity?
  • 2D Scans
  • Single Robot


Presenter Notes
Mostly 2d scans



Single robot

8
What do we want?
  • Incremental
  • Real-time
  • Multi-Robot
  • 3D Scans
  • Handles cycles
  • Low Complexity


Presenter Notes
Thats what is needed..

Low complexity in terms of the 3d point cloud..
Soo the software can run on low end comsumer
laptops

9
Approach
  • Combine EM (posterior estimation) with
    incremental map construction
  • Makes it real-time.
  • Allows backwards correction
  • Allows multi-robot operations.
  • Generate 3d Point Cloud
  • Robust


Presenter Notes
What is desired!

10
Overview
  • Map is collection of scans and pose
  • FIND
  • Most likely map given data


Presenter Notes
Data is collection of laser scans and odometry
(or scan matching estimates)

11
Likelihood function
  • Motion Probability Model
  • Perceptual/Sensor Model


Presenter Notes


Perception model inherited from documentation
about scan matching

12
Likelihood function

Presenter Notes
Markov localisation

13
Likelihood function

Presenter Notes
Occupancy grids

14
Incremental
  • Find scan and pose
  • Add to map
  • Forget about it
  • No backwards correction
  • Pose errors can grow large


Presenter Notes
Give example from robotics class

15
Incremental

Presenter Notes
St pose

Ot laser

At odometry



16
Incremental

Presenter Notes
ERROR

17
Fun Time
  • Incremental Mapping with Posteriors


Presenter Notes
Dt laser scans and odometry

Mt map

St Pose

Probability distribution over poses given past
sensor datas

18
Monte Carlo
  • Markov Localization


Presenter Notes
Assume equal distribution (probability of
location). Update it using the sensor model

Here we assume we know where we start from.

Use Markov Localization approach for static maps.




basically an implementation of the particle filter

19
Monte Carlo

Presenter Notes
Imp sampling.. resampling

20
Monte Carlo

Presenter Notes
http//robots.stanford.edu/movies/sca80a0.avi

21
Backwards Correction
  • No loop closing -gt
  • When loop closed -gt
  • backwards correction


Presenter Notes
St is incremental.. Other is incremental with
belief

22
Backwards Correction
  • How many scans to fix?
  • Distribute the error over the scans
  • Gradient search (check all possible poses)


Presenter Notes
Scans involved in the loop closing



Gradient search checks all poses from the maximum
likelihood



Extremely robust approach Fast (between two
sensor measuements

23
Multi-Robot
  • Use posterior estimation
  • Assumption


Presenter Notes
Assumption (all robots in the same map of leader
robot)

24
Multi-Robot
  • Multi-Robot


Presenter Notes
Assumption (all robots in the same map of leader
robot)

25
Multi-Robot

Presenter Notes
http//robots.stanford.edu/movies/Map4b.avi

26
3d Scanning
  • With 2D scanner?
  • Naïve solution
  • Now we simplify


Presenter Notes
Connect nearby lasers into polygons Noise Too
Complex

The filter works by not taking into account two
measurements are larger than the expected
measurement (due to motion)

Fuse polygons which look similar when rendered.



27
3d simplification
  • Filter out outlier
  • Constraints
  • Reduce Polygons


Presenter Notes
Connect nearby lasers into polygons Noise Too
Complex

The filter works by not taking into account two
measurements are larger than the expected
measurement (due to motion)

Fuse polygons which look similar when rendered.



28
Results
  • Characteristics
  • Scan added to map if robot moves 2 meters
  • All scans used in localization
  • Random error in odometer


Presenter Notes
Real-time on a low-end PC

29
Results

Presenter Notes


http//robots.stanford.edu/movies/mapping1-new.avi

30
Results
  • Results without odometry


Presenter Notes
However only as good as scan matching

No features would fail.

Featureless corridor



EXPLAIN SCAN MATCHING!!!!!!!!

31
Results
  • Rough terrain


Presenter Notes
The robot had week odometry depending on the
surface

Environment not fully orthogonal

32
Results
  • Multi-Robot
  • 3d Mapping

33
Results

Presenter Notes
http//robots.stanford.edu/movies/wean.mpg

34
Conclusion
  • Incremental method
  • SLAM
  • 2d laser range finders
  • Scan-Matching


Presenter Notes
incremental method for concurrent mapping and
localization for mobile robots equipped with 2D
laser range finders. The approach uses a fast
implementation of scan-matching for mapping,

35
Conclusion
  • Sample based Probabilistic method
  • Dual Laser system
  • Compact 3d Maps
  • Real-time


Presenter Notes
paired with a sample-based probabilistic method
for localization. Compact 3D maps are generated
using a multi-resolution approach adopted from
the computer graphics literature, fed by data
from a dual laser system. Our approach builds 3D
maps of large, cyclic environments in real-time.

36
Conclusion
  • Large Cyclic environment
  • Very Robust
  • Absence of Odometry
  • Award winning


Presenter Notes
It is remarkably robust.

Experimental results illustrate that accurate
maps of large, cyclic environments can be
generated even in the absence of any odometric
data.

This paper won the Best Conference paper award at
the IEEE International Conference on Robotics and
Automation, 2000 in San Francisco.

37
Finally
About PowerShow.com