Vision Based Control

jabgoldfishAI and Robotics

Oct 19, 2013 (3 years and 5 months ago)

124 views

Vision Based Control
Motion

Matt Baker

Kevin VanDyke

Robots


Today’s robots perform complex tasks with
amazing precision and speed


Why then have they not moved from the
structure of the factory floor into the “real”
world? What is the limiting factor?

Vision

A “Seeing Robot”


A robot that can perceive
and react in complex and
unpredictable surroundings


This is not possible with the
marker
-
based systems in
use in most laboratory
vision
-
based control
systems

Common reasons for failure of
vision systems


Small changes in the environment can
result in significant variations in image
data


Changes in contrast


Unexpected occlusion of features

Robustness


Stable measurements of
local feature attributes,
despite significant
changes in the image
data, that result from
small changes in the 3D
environment
[1].

Enhanced Techniques


The Hough
-
Transform


Robust color classification


Occlusion prediction


Multisensory visual servoing

Hough Transform


Used to extract geometrical object features from
digital images

Hough Transform (con’t)


Features are extracted by detecting
maximums in the image


Example geometric features encountered:

Lines:

Circles:

Ellipses:

Hough Transform (cont’d)


Advantages


Noise and background clutter do not impair
detection of local maxima


Partial occlusion and varying contrast are
minimized


Negatives


Requires time and space storage that
increases exponentially with the dimensions
of the parameter space

Hough Transform (con’t)


a real
-
time application of HT requires both a fast
image preprocessing step and an efficient
implementation

Implementation of a circle tracking algorithm based on HT

Robust color classification


Color has high disambiguity power


Real
-
time is required


Supervised color segmentation


The color distribution of the current scene is
analyzed and colors that do not appear in the
scene are used as marker colors


These markers are then used as the input to the
visual servoing system


Colors represented by their hue
-
saturation value
(H&S relate to color, V relates to brightness)



Robust color classification (con’t)


Color segmentation


Choose four colors as
marker colors


Color markers brought
onto object we wish to
track



markers outlined


Color distribution
computed


Initial segmentation

Model
-
based handling of occlusion


The previous two techniques take care of
bad illumination and partial occlusion


What about aspect changes (complete
occlusion)?


Build and maintain a 3D model of the
observed objects so they can be tracked
despite occlusion


Then use prediction

Tracking system model

Sensor data

Feature
extraction

3D pose
estimation

Robot
control

Pose
prediction

Visibility
determination

Feature
selection

Geometric
model

Designed to handle aspect changes online

Prediction


Extract measurements of object features based on raw
sensor data


Estimate the spatial position and orientation of the target
object


Based on history of estimated poses and assumptions
about the object motion you can predict an object pose
expected in next sampling interval


With predicted pose and 3D model we are able to
determine feature visibility in advance


Guide the feature extraction process for the next frame
without the risk of searching for occluded features

Model
-
based handling of occlusion
(con’t)


Efficient Hidden Line Removal


Explicit modeling of curved object structures allows us
to remove
virtual lines



or lines that do not have a
physical correspondence in the camera image

Object tracking with visibility
determination

Multisensory Servoing


Redundant information is used to increase the
performance of the servoing system as well as
the robustness against failing sensors

Vision Controlled Robot Model

Conclusions


We explored a variety of
image processing
techniques that can
significantly improve the
robustness of visual
servoing systems


These techniques can be
implemented in modern
robot vision control
systems


Techniques such as
these will make machine
vision in robots a reality in
the near future