SPIE'01CIRL-JHU1 Dynamic Composition of Tracking Primitives for Interactive Vision-Guided Navigation D. Burschka and G. Hager Computational Interaction.
Post on 29-Dec-2015
214 Views
Preview:
Transcript
SPIE'01 CIRL-JHU 1
Dynamic Composition of Tracking Primitives for Interactive
Vision-Guided Navigation
D. Burschka and G. Hager
Computational Interaction
and
Robotics Laboratory (CIRL)
Johns Hopkins University
SPIE'01 CIRL-JHU 2
Outline
Introduction Motivation – Navigation Strategies
Tracking-System Architecture Pre-Processing New Tracking Definition Feature Identification
Results Conclusions
SPIE'01 CIRL-JHU 3
Navigation Strategies
Sensor-Based Control control signals for the robot are generated directly from the visual input
i i 1
Map-Based Navigation pre-processed sensor data is stored in a geometrical representation of the envi- ronment (map). Path plan- ning+strategy algorithms are used to define the actions of the robot
SPIE'01 CIRL-JHU 4
Tracking Primitives
Dynamic Vision(XVision)
algorithms
Color Tracking Pattern Tracking Disparity tracking
SPIE'01 CIRL-JHU 5
XVision as Tracking Tool
Dynamic Vision(XVision)
algorithms
applications
SPIE'01 CIRL-JHU 6
Tracking-System Architecture
Templates(SSD)
Hue(Color Blob)
Disparity(Disparity Region)
Points Curves
Feature Extraction
Feature-Based
Region-Based
Domain Conversion
Tracking Module
User/Task
Physical Hardware-Layer
Image Processing-Layer
Tracking-Layer
Coordination-Layer
Fe
atu
re I
de
nti
fic
ati
on
La
ye
r
SPIE'01 CIRL-JHU 7
Dynamic Composition of Tracking Cues
SPIE'01 CIRL-JHU 8
Tracking-System Architecture
Templates(SSD)
Hue(Color Blob)
Disparity(Disparity Region)
Points Curves
Feature Extraction
Feature-Based
Region-Based
Domain Conversion
Tracking Module
User/Task
Physical Hardware-Layer
Image Processing-Layer
Tracking-Layer
Coordination-Layer
Fe
atu
re I
de
nti
fic
ati
on
La
ye
r
SPIE'01 CIRL-JHU 9
Segmentation in the ColorSpace
- HSI representation of color space
- Variable resolution gridding of space
Intensity
Hue
Saturation
SPIE'01 CIRL-JHU 10
Segmentation in the Disparity Domain
SPIE'01 CIRL-JHU 11
Tracking-System Architecture
Templates(SSD)
Hue(Color Blob)
Disparity(Disparity Region)
Points Curves
Feature Extraction
Feature-Based
Region-Based
Domain Conversion
Tracking Module
User/Task
Physical Hardware-Layer
Image Processing-Layer
Tracking-Layer
Coordination-Layer
Fe
atu
re I
de
nti
fic
ati
on
La
ye
r
SPIE'01 CIRL-JHU 12
State Transitions in the Tracking Process
SPIE'01 CIRL-JHU 13
State Information saved in the Tracking Module
Information about the object in the real scene is shared between the different Image Identifications:
Position in the imageSize of the regionRange in the current image domainShape ratio in the imageCompactness of the region
SPIE'01 CIRL-JHU 14
Tracking-System Architecture
Templates(SSD)
Hue(Color Blob)
Disparity(Disparity Region)
Points Curves
Feature Extraction
Feature-Based
Region-Based
Domain Conversion
Tracking Module
User/Task
Physical Hardware-Layer
Image Processing-Layer
Tracking-Layer
Coordination-Layer
Fe
atu
re I
de
nti
fic
ati
on
La
ye
r
SPIE'01 CIRL-JHU 15
Quality Value for Initial Search
cd 10
diD
,1min
iDiCi ,min
R
corriC A
A
SPIE'01 CIRL-JHU 16
Problem in the Disparity Domain
SPIE'01 CIRL-JHU 17
Ground Plane Suppression
SPIE'01 CIRL-JHU 18
Results Obstacle Detection
SPIE'01 CIRL-JHU 19
Results Dynamic Composition
SPIE'01 CIRL-JHU 20
Conclusions and Future Work:
Dynamic Composition of the two Basic Feature Identification tools allowed robust initial selection and navigation through a door
Extension to the entire set of Feature Identification tools is our next step
The developed algorithms allow robust obstacle avoidance
SPIE'01 CIRL-JHU 21
Additional Information:
Web: http://www.cs.jhu.edu/CIRL http://www.cs.jhu.edu/~burschka
top related