Presented by Adewole Ayoade Computer Vision Final Project Presentation April 30, 2014

advertisement
Presented by
Adewole Ayoade
Computer Vision Final Project Presentation
April 30, 2014
•
•
•
•
•
Introduction
Previous Works
Methodology
Experiment and Results
Conclusion
Problem Statement
•Aiding robot localization
GPS failure (Multi-path signals)
Compass unreliable (Magnetic variation)
Dead reckoning (drifting over time)
Why Fiducial Marker System
•
•
•
•
•
Structured Environment
Existing Map
Computationally less expensive
Non-unique future environment
Inexpensive to deploy
Objectives
•
•
•
•
Design a robust fiducial marker system
Evaluate the marker system
Compare with other existing marker systems
Use the system for robot localization
Application
• Augmented Reality
• Inspection Robot
Existing Marker System
• ARToolKit
Binarization of input image
Uses orthogonal templates
• ARToolkitPlus
Digitally encoded payload
• ARTag
Detection mechanism based on
image gradient, robust to lighting
variation
• AprilTag
Quad detectors
They were developed for indoor use
Design
13.66cm
Origin
1
11.40 cm
2
3
14.25 cm
4
Made from Contrasting Concentric Circles
Payload Region
Detection
•
•
•
•
•
•
•
•
•
•
•
•
Read in image
Convert to grayscale
Perform local thresholding
Close image with 2X2 disk
Find image complement
Perform blob analysis on both
Locate objects in both images with very close
centroids
Find ratio of black area to white area
Pick candidate features whose ratio is between one
and six
No detection if remaining candidate points are less
than four
If greater than four, find four points with smallest
average distance between all pairs of candidate points
Proceed to find correspondence if four points
Finding Correspondence
• Find the black blob with the biggest area
(point 1)
• Find the slope of the other three points with
it
• Find the average difference between slopes
and the largest is point 4
• The farthest of the remaining candidate is
point 3
• The remainder is point 2
Identification
• Find the Line parallel
to slope between
point 1 and point 3
• Starting from point 4,
find number of pixels
with their state values
• Calculate their ratio
• If sum of ratio is not
10, it is invalid
• From right to left,
pick the last 7 bands
and they form the
code
2
1
3
4
Payload
Region(7 bands)
Start bands(3
bands)
Pose Estimation
• Using calibrated camera
• Least square pose estimation
E=|f(x)-y|^2
x = unknown pose parameters
f (x) = predicted image points
y = actual observed image points
• Obtain pose of the fiducial(Model) with respect
to the Camera CTM
Localization and Extended Kalman Filter
• Camera pose relative to the vehicle (V) is known
CT = [VT ]-1
V
C
MT = [CT ]-1 known from pose
C
M
estimation
WT
Model is known relative to
M
World from the Map
WT
V
= WTM X MTC X CTV
Obtain X,Y and Heading from the Transformation
Matrix which is fed into Extended Kalman Filter
as a measurement update.
Evaluation and Result
• Outdoor Environment
• Afternoon and Evening for light Variation
• Dirty Targets vs Clean Targets
Evaluation and Result
Day Time Tag Type
Afternoon Clean Tag
Dirty Tag
Very Dirty
Tag
Maximum Distance(m)
Angle Detected
Identified
0
12.3
10
45
10
6.5
60
5
70
5.5
0
12.3
10
0
0
45
40
10
12.3
5
4
4
5
4
4
Evaluation and Result
Maximum Distance(m)
Day Time Tag Type
Evening
clean
Verydirty
Angle Detected
Identified
0
7.8
7.5
45
4.7
4.7
0
1.3
1.3
45
1.3
1.3
Robot Localization
• Created a local
coordinate system (50m
by 15m)
• Placed 13 different
fiducials especially at
corner points
• Robot driven with
remote control
• With and without
fiducial integration
2nd Floor Plan, Building 3, Petroleum
Institute
Robot Localization Result
Without fiducial
With fiducial
• Fiducial Marker System can aid robot localization
• Dirty tags can be problematic
Download