8/14/2019 Lecture5 Vision Sensor Part2
1/17
Depth from Focus (1)
L diameter of the lens or
aperture
Image formation formula:R - the radius of the blur circle on
the image plane
8/14/2019 Lecture5 Vision Sensor Part2
2/17
Depth from Focus (2)
Measure of sub-image intensity (I) gradient:
Limitation of depth from focus
techniques:(1) They lose sensitivity as
objects move farther away (given
a fixed focal length)
(2) Slow focusing
Sharpness change
between the near
surface and far
surface
The lens optics are actively searched in order to maximize focus. Not for mobile robots
due to its slow focusing.
8/14/2019 Lecture5 Vision Sensor Part2
3/17
8/14/2019 Lecture5 Vision Sensor Part2
4/17
Depth from Defocus
The second equation relates the depth of scene points viaR to the
observed image g. SolvingR would provide us the depth.
Another unknownf(x, y) the focused image which can be obtainedby a pinhole aperture lens model.
In summary, the basic advantage of the depth from defocus method is
its extremely fast speedNo correlation search problem
No need to capture scene at different perspectives, which may lead to
occlusions and the disappearance of objects in a second view
Disadvantage: accuracy decreases with distance, as with all visual
methods for ranging.
8/14/2019 Lecture5 Vision Sensor Part2
5/17
Stereo Vision
Idealized camera geometry for stereo vision
Disparity between two images -> Computing of depth
From the figure it can be seen that
8/14/2019 Lecture5 Vision Sensor Part2
6/17
Stereo Vision
1. Distance is inversely proportional to disparity
closer objects can be measured more accurately
2. Disparity is proportional to b.
For a given disparity error, the accuracy of the depth estimate
increases with increasing baseline b.
However, as b is increased, some objects may appear in one camera,
but not in the other.
3. A point visible from both cameras produces a conjugate pair.
Conjugate pairs lie on epipolar line (parallel to the x-axis for the
arrangement in the figure above)
8/14/2019 Lecture5 Vision Sensor Part2
7/17
Stereo Vision the general case
The same point P is measured differently in the left camera image :
where
R is a 3 x 3 rotation matrix
r0 = offset translation matrix
The above equations have two uses:
We can find rrif we knew R and rl and r0. Note:For perfectly aligned cameras R=I (unity matrix)
We can calibrate the system and find r11, r12 given corresponding values of xl, yl, zl, xr, yrand zr.
We have 12 unknowns and require 12 equations:
we require 4 conjugate points for a complete calibration.
Note: Additionally there is a optical distortion of the image
left camera
coordinate system
P
yl
xl
zl
yr
x
zr
rl
rr
right camera
coordinate system
8/14/2019 Lecture5 Vision Sensor Part2
8/17
Stereo Vision the general case
Suppose the calibration is complete
We know the pixels ofP on the image planes of each camera (xl, yl, zl), and (xr, yr,
zr). Given the focal lengthfof the camera, we have
'
'and
'
'
l
ll
l
ll
z
y
f
y
z
x
f
x==
'')(
'')(
'')(
03333231
02232221
01131211
rl
ll
rr
lll
rr
lll
zrzrf
y
rf
x
r
zf
yrzr
f
yr
f
xr
zf
xrzr
f
yr
f
xr
=+++
=+++
=+++
The same process can be
used to identify values forx
andy
Now we want to recover 'and' rl zz
8/14/2019 Lecture5 Vision Sensor Part2
9/17
Stereo Vision
Calculation of Depth
The key problem in stereo is now how do we solve the correspondenceproblem?
Gray-Level Matching match gray-level wave forms on corresponding epipolar lines
brightness = image irradiance I(x,y)
Zero Crossing of Laplacian of Gaussian is a widely used approach foridentifying feature in the left and right image
8/14/2019 Lecture5 Vision Sensor Part2
10/17
Zero Crossing of Laplacian of Gaussian
Identification of features that are stable and match well
Laplacian of intensity image
Convolution with P:
Step / Edge Detection
in Noisy Image
filtering through
Gaussian smoothing
P=
8/14/2019 Lecture5 Vision Sensor Part2
11/17
Stereo Vision Example
Extracting depth information from a stereo image
a1 and a2: left and right image
b1 and b2: vertical edge filtered
left and right image;filter = [1 2 4 -2 -10 -2 4 2 1]
c: confidence image:
bright = high confidence (good texture)
d: depth image:
bright = close; dark = far
8/14/2019 Lecture5 Vision Sensor Part2
12/17
SVM Stereo Head Mounted on an All-terrain Robot
Stereo Camera
Vider Desing
www.videredesign.com Robot
Shrimp, EPFL
Application of Stereo Vision Traversability calculation based on
stereo images for outdoor navigation
Motion tracking
8/14/2019 Lecture5 Vision Sensor Part2
13/17
Color Tracking Sensors
Color represents an environmental characteristic that is orthogonal to
range, and it represents both a natural cue and an artificial cue that can
provide new information to a mobile robot
Advantages
Detection of color is straightforward
It can combine (sensor fusion) with existing cues, such as range findings,to have significant information gains
8/14/2019 Lecture5 Vision Sensor Part2
14/17
Color Tracking Sensors
Motion estimation of ball and robot for soccer playing using color
tracking
8/14/2019 Lecture5 Vision Sensor Part2
15/17
Color Tracking Sensors
CMUcam robotic vision sensor
http://www.cs.cmu.edu/~cmucam/gallery.html
CMOS imaging sensor and high-speed microprocessors at 50+ Mhz range
An external processor configures the sensors streaming data mode, such asspecifying tracking mode for a bounded YUV value set.
The YUVmodel defines a color space in terms of one luminance (Y) and twochrominance (U and V) components. .
The vision sensor processes the data in real-time and outputs high-levelinformation to the external consumer.
http://www.cs.cmu.edu/~cmucam/gallery.htmlhttp://www.cs.cmu.edu/~cmucam/gallery.html8/14/2019 Lecture5 Vision Sensor Part2
16/17
Homework #4
Has been posted on Stevens Pipeline
Prepare your project proposal and present your proposal on Oct. 15.
The mid-term exam grade will be based on your presentation of theproposal
8/14/2019 Lecture5 Vision Sensor Part2
17/17
What should you include in your proposal?
Introduction: what kind of problem you want to solve? Why this problem
is important for the mobile robot? What other people have done for this
problem (list several reference papers)?
Approach: How are you going to solve this problem? You can either pick
one approach which is available from papers and improve it at some
levels, or propose some new ideas. (Extra credit will be given to the new
ideas).At this level, you dont have to specify in details how to tackle the
problem, only indicate which method you will focus on to improve, or
what kind of draft idea you may want to pursuit.