Lecture6 PDF

Autonomous Mobile Robots
Perception
Sensors
Vision
Uncertainties, Line extraction from laser scans
Zürich Autonomous Systems Lab

Lecture 6 - Perception - Vision
Lec. 6
2
Image Intensities & Data reduction
 Monochrome image  matrix of intensity values
 Typical sizes:
  320 x 240 (QVGA)
  640 x 480 (VGA)
  1280 x 720 (HD)
 Intensities sampled to 256 grey levels  8 bits
 Images capture a lot of information
Useful
cues
Heavy
Processing
 Reduce the amount of input data:

preserving useful info & discarding redundant info
© R. Siegwart , D. Scaramuzza and M.Chli, ETH Zurich - ASL

Lec. 6
3
What is USEFUL, What is REDUNDANT ?
218 218 218 218 218 218 218 218 218 218 218
218 218 218 218 218 218 218 218 218 218 218
218 218 218 218 218 218 218 218 218 218 218
218 218 218 218 218 218 218 218 218 218 218
218 218 218 218 218 218 218 218 218 218 218
218 218 218 218 218 218 218 218 218 218 218
218 218 218 218 218 218 218 218 218 218 218
218 218 218 218 218 218 218 218 218 218 218
218 218 218 218 218 218 218 218 218 218 218
218 218 218 218 218 218 218 218 218 218 218
218 218 218 218 218 218 218 218 218 218 218
Image from http://www.flickr.com/photos/mukluk/241256203/

Lec. 6
4
208 208 208 208 208 208 208 208 208 208 208
208 208 208 208 208 208 208 208 208 208 208
208 208 208 208 208 208 208 208 208 208 208
208 208 208 208 208 208 208 208 208 208 208
208 208 208 208 208 208 208 208 208 208 208
208 207 208 208 208 208 208 208 208 208 208
0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0

Lec. 6
5
229 229 229 229 229 229 229 229 229 229 229
229 229 229 229 229 229 229 229 229 229 229
229 229 229 229 229 229 229 229 229 229 229
229 229 229 229 229 229 229 229 229 229 229
229 229 229 229 229 230 229 229 229 229 229
5 17 31 7 1 0 229 229 229 229 229
0 0 1 0 0 0 229 229 229 229 229
0 0 0 0 0 0 229 229 229 229 229
0 0 0 0 1 4 229 229 229 229 229
0 0 0 0 0 11 229 229 229 229 229
0 0 0 0 0 5 229 229 229 229 229

Lec. 6
6
Today’s Topics
Sections 4.3 – 4.5 of the book
 Image Filtering
  Correlation
  Convolution
 Edge / Corner Detection
 Image Features
  Harris corners
  SIFT features

Lec. 6
7
Image filtering
  “filtering”: accept / reject certain components
  Example: a lowpass filter allows low frequencies
 blurring (smoothing) effect on an image – used to reduce image noise
  Smoothing can also be achieved by spatial filters instead of frequency filters.
Lowpass filtered image Highpass filtered image

Lec. 6
8
Spatial filters
  Sxy : neighborhood of pixels around the point (x,y) in an image I
  Spatial filtering operates on Sxy to generate a new value for the corresponding pixel at
output image J
Sxy
2M+1
2N+1
Image Filtered Image
  For example, an averaging filter is:

Lec. 6
9
Linear, Shift-invariant filters
  Linear: every pixel is replaced by a linear combination of its neighbours
  Shift-invariant: the same operation is performed on every point on the image

  Basic & very useful filtering operations:
  Correlation
  Convolution
  Brief study of these filters in the simplest case of 1D images (i.e. row of pixels)
& their extension to 2D
A more detailed tutorial on this here: http://www.cs.umd.edu/~djacobs/CMSC426/Convolution.pdf

Lec. 6
10
Correlation
 An averaging filter
… 5 4 2 3 7 4 6 3 6 6 …
× × ×
1/3 1/3 1/3
=
=
4/3 2/3 3/3
9/3
 How to handle boundaries?

  Ignore filtered values at boundaries
  Pad with zeros
  Pad with first/last image values

Lec. 6
11
Correlation
 An averaging filter
… 5 4 2 3 7 4 6 3 6 6 …
× × ×
1/3 1/3 1/3 Filter, kernel, mask, window
=
2/3 3/3 7/3
9/3 12/3
 Formally, Correlation is
Other examples of
 In this smoothing example
smoothing filters?

Lec. 6
12
Constructing Filter from a Continuous Fn
 Common practice for image smoothing: use a Gaussian
: controls the amount of smoothing
Normalize filter so that values always add up to 1
 Near-by pixels have a bigger influence on the averaged value rather than
more distant ones

Lec. 6
13
Constructing Filter from a Continuous Fn
 Common practice for image smoothing: use a Gaussian
99% : controls the amount of smoothing
Normalize filter so that values always add up to 1
 Near-by pixels have a bigger influence on the averaged value rather than
more distant ones

Lec. 6
14
Taking derivatives with Correlation
 Derivative of an image: quantifies how quickly intensities change

(along the direction of the derivative)
 Approximate a derivative operator:
× × ×
-1/2 0 1/2

Lec. 6
15
Matching using Correlation
 Find locations in an image that are similar to a template
 Filter = template 3 8 3  test it against all image locations
3 2 4 1 3 8 4 0 3 8 7 7
 Similarity measure: Sum of Squared Differences (SSD)
26 37 21 50 54 1 50 65 59 16 42 17
Ignoring the filter values beyond the image-boundaries,

this value is computed as (8-3)2+(3-4)2 = 26

Lec. 6
16
Matching using Correlation
3 2 4 1 3 8 4 0 3 8 7 7
 Similarity measure: Sum of Squared Differences (SSD)
Correlation
26 37 21 50 54 1 50 65 59 16 42 17
 Similarity measure: Correlation?

30 37 41 29 51 85 56 21 48 86 101 77

Lec. 6
17
NCC: Normalized Cross Correlation
3 2 4 1 3 8 4 0 3 8 7 7
 Correlation value is affected by the magnitude of intensities

 Similarity measure: Normalized Cross Correlation (NCC)
0.919 0.759 0.988 0.628 0.655 0.994 0.691 0.464 0.620 0.860 0.876 0.859

Lec. 6
18
ZNCC: Zero-mean Normalized Cross Correlation
3 2 4 1 3 8 4 0 3 8 7 7
 For better invariance to intensity changes

Similarity measure: Zero-mean Normalized Cross Correlation (ZNCC)
where

Lec. 6
19
Correlation as an inner product
 Considering the filter F and the portion of the image Ix as vectors then
their correlation is:
 In NCC and ZNCC we considering the unit vectors of F and Ix , hence we
measure their similarity based on the angle

Lec. 6
20
Correlation in 2D
“separable” filter
 Example: Constant averaging filter
 If i.e. this is a square filter
 2D Correlation  no. multiplications per pixel =

no. additions per pixel =
This example was generated with a 21x21 mask

Lec. 6
21
Correlation in 2D

 2 × 1D Correlation  no. multiplications per pixel =

Lec. 6
22
Correlation in 2D

(with const. factor) no. additions per pixel =

Lec. 6
23
2D Gaussian Smoothing
  A general, 2D Gaussian
  We usually want to smooth by the same amount in both x and y directions
  So this simplifies to:
  Another separable filter

Lec. 6
24
Convolution
  Convolution is equivalent to: Correlation with a flipped filter before correlating
  CONVOLUTION So if
  CORRELATION Then,
  Likewise, in 2D we flip the filter both horizontally & vertically
  Key difference between correlation and convolution is that convolution is associative:
  Very useful!
  Example: smooth an image & take its derivative  convolve the Derivative filter with the
Gaussian Filter & convolve the resulting filter with the image

Lec. 6
25
Edge Detection
  Ultimate goal of edge detection: an idealized line drawing.
  Edge contours in the image correspond to important scene contours.

Lec. 6
26
Edge = intensity discontinuity in one direction
  Edges correspond to sharp changes of intensity
  Change is measured by 1st order derivative in 1D
  Big intensity change  magnitude of derivative is large
  Or 2nd order derivative is zero.

Lec. 6
27
1D Edge detection
  Image intensity shows an obvious change
  Where is the edge?  image noise cannot be ignored

Lec. 6
28
Solution: smooth first
  Where is the edge? Edges occur at maxima/minima of

Lec. 6
29
Derivative theorem of convolution
  .
  This saves us one operation:
Edges occur at maxima/minima of

Lec. 6
30
Zero-crossings
  Locations of Maxima/minima in are equivalent to zero
-crossings in
: Laplacian of Gaussian operator
Edges occur at zero-crossings of

Lec. 6
31
2D Edge detection
  Find gradient of smoothed image in both directions
Usually use a separable
filter such that:
  Discard pixels with (i.e. edge strength), below a certain below

a certain threshold
  Non-maximal suppression: identify local maxima of

 detected edges

Lec. 6
32
2D Edge detection: Example
: original image (Lena image)

Lec. 6
33
: Edge strength
Lec. 6
34
Thresholding
Lec. 6
35
Thinning: non-maximal suppression

Image Feature
Extraction
•  Edges (seen before)
•  Points:
o  Harris corners
o  SIFT features

Lec. 6
37
Example: Build a Panorama
This panorama was generated using AUTOSTITCH (freeware)
Build your own: http://www.cs.ubc.ca/~mbrown/autostitch/autostitch.html

Lec. 6
38

Lec. 6
39

Lec. 6
40
How to build a panorama?
  We need to match (align) images

Lec. 6
41
 Detect feature points in both images

Lec. 6
42
 Find corresponding pairs

Lec. 6
43
 Find corresponding pairs
 Use these pairs to align images

Lec. 6
44
Matching with Features
  Problem 1:
  Detect the same points independently in both images,
if they are in the field of view
no chance to match!
We need a repeatable detector

Lec. 6
45
Matching with Features
  Problem 2:
  For each point, identify its correct correspondence in the other image(s)
We need a reliable and distinctive descriptor

Lec. 6
46
More motivation…
  Feature points are used also for:
  Robot navigation (we will see in tomorrow’s exercise)
  Object recognition (live demo)
  Image alignment (panoramas)
  3D reconstruction
  Motion tracking
  Indexing and database retrieval -> Google Images or http://tineye.com
  … other

Harris Corner Detector
C.Harris, M.Stephens. “A Combined Corner and Edge Detector”. 1988

Lec. 6
48
Identifying Corners
• Key property: in the region around a corner, image gradient has

two or more dominant directions
• Corners are repeatable and distinctive

Lec. 6
49
Identifying Corners
• Shifting a window in any direction should give a large

change in intensity in at least 2 directions
flat region:  edge :  corner : 

no intensity no change along the significant change in
change edge direction at least 2 directions

Lec. 6
50
How do we implement this?
  Two image patches of size P one centered at and one centered at
The Sum of Squared Differences between them is:
  Let and . Approximating with a 1st order Taylor expansion:
  This produces the approximation
  Which can be written in a matrix form as

Lec. 6
51
  M is the “second moment matrix”
  Since M is symmetric, we can rewrite M as
where λ1 and λ2 are the eingenvalues of M

Lec. 6
52
 The Harris detector analyses the eigenvalues of M to decide if we are in
presence of a corner or not
 i.e. looks for large intensity changes in at least 2 directions
 Visualize M as an ellipse with axis-lengths determined by the eigenvalues

and orientation determined by R

Lec. 6
53
Visualization of second moment matrices

Lec. 6
54
Visualization of second moment matrices

Lec. 6
55
Corner response function
Does the patch describe a corner or not?
  No structure:
SSD is almost constant in all directions, so it’s a flat region
  1D structure: is large (or vice versa)

SSD has a large variation only in one direction, which is the one perpendicular to the
edge.
  2D structure: are both large

SSD has large variations in all directions and then we are in presence of a corner.
Computation of eigenvalues is expensive  Harris and Stephens suggested using a

“cornerness function” instead:
Where k is a between 0.04 and 0.15
  Last step of Harris corner detector: extract local maxima of the cornerness function

Lec. 6
56
Harris Detector: Workflow

Lec. 6
57
 Compute corner response C

Lec. 6
58
 Find points with large corner response: C > threshold

Lec. 6
59
 Take only the points of local maxima of thresholded C

Lec. 6
60

Lec. 6
61
Harris detector: properties
  How does the Harris detector behave to common image

transformations?
  Will we be able to re-detect the same image patches (Harris corners)
when the image exhibits changes in
  Rotation,
  View-point,
  Zoom,
  Illumination?
  Identify properties of detector & adapt accordingly

Lec. 6
62
Harris Detector: Some Properties
  Rotation invariance
image 1 image 2
Ellipse rotates but its shape (i.e. eigenvalues) remains the same
Corner response C is invariant to image rotation

Lec. 6
63
  But: non-invariant to image scale!
image 1 image 2
All points will be Corner!

classified as edges

Lec. 6
64
  Quality of Harris detector for different scale changes
Repeatability:
# correspondences detected Scaling the image by ×2

 ~20% of correspondences get
# correspondences present detected

Lec. 6
65
Summary on Harris properties
 Harris detector: probably the most widely used & known corner detector
 The detection is Invariant to

  Rotation
  Linear intensity changes
  note: to make the matching invariant to these we need a suitable detector and
matching criterion (e.g. SSD on patches is not rotation- or affine- invariant)
 The detection is NOT invariant to

  Scale changes
 Geometric affine changes: an image transformation which distorts the

neighborhood of the corner, can distort its ‘cornerness’ response

Lec. 6
66
Scale Invariant Detection
  Consider regions (e.g. discs) of different sizes around a point
  aim: corresponding regions look the same in image space, when the
appropriate scale-change is applied

Lec. 6
67
  The problem: how do we choose corresponding regions (discs)

independently in each image?

Lec. 6
68
  Approach: design a function to apply on the region (disc), which is “scale invariant”
(i.e. remains constant for corresponding regions, even if they are at different scales)
example: average image intensity over corresponding regions

(even of different sizes) should remain constant
  Average intensity value enclosed in each disc, as a function of the disc-radius:

I I Image 2
Image 1
scale = 1/2
region size region size

Lec. 6
69
Identify the local maximum in each response

 these occur at corresponding region sizes
The corresponding scale-invariant region size is found

in each image independently!
I I Image 2
Image 1
scale = 1/2
s1 region size s2 region size

Lec. 6
70
 A “good” function for scale detection: has one clear, sharp peak
I I I Good !
bad bad
region size region size region size
 Sharp, local intensity changes in an image, are good regions to monitor for
identifying relative scale in usual images.

Lec. 6
71
  Functions for determining scale:

convolve image with kernel for identifying sharp intensity discontinuities
Kernels:
(Laplacian of Gaussian)
(Difference of Gaussians)
Note: These kernels are invariant to
scale and rotation

Lec. 6
72
LoG for Scale invariant detection
  Response of LoG for corresponding regions
  “Harris-Laplacian” multi-scale detector

Lec. 6
73
Scale Invariant Detectors
  Experimental evaluation of detectors

w.r.t. scale change
Repeatability:
# correspondences detected
# correspondences present

Lec. 6
74
SIFT features
 SIFT: Scale Invariant Feature Transform

an approach for detecting and describing regions of interest in an image.
Developed by D. Lowe.
 SIFT features are reasonably invariant to changes in: Rotation, scaling,

small changes in viewpoint, illumination
 Very powerful in capturing + describing distinctive structure, but also

computationally demanding
Main SIFT stages:

1.  Extract keypoints + scale
2.  Assign keypoint orientation
3.  Generate keypoint descriptor

Lec. 6
75
SIFT detector (keypoint location + scale)
1.  Scale-space pyramid: subsample and blur original image
2.  Difference of Gaussians (DoG) pyramid: subtract successive smoothed images
Blur
3.  Keypoints: local extrema in the DoG
pyramid
Subsample
DoG:

Lec. 6
76
SIFT: assign keypoint orientation
Find ‘orientation’ of keypoint to achieve rotation

invariance
  Sample intensities around the keypoint
  Compute a histogram of orientations of intensity
gradients
  Peaks in histogram: dominant orientations

  Keypoint orientation = histogram peak
  If there are multiple candidate peaks, construct a

different keypoint for each such orientation

Lec. 6
77
SIFT descriptor
  Descriptor : “identy card” of keypoint
  Simplest descriptor: matrix of intensity values around a keypoint (image patch)
  Ideally, the descriptor is highly distinctive: allows recognition of a given feature uniquely
among many others!
  SIFT descriptor: 128-long vector

  Describe all gradient orientations relative to the Keypoint Orientation
  Divide keypoint neighbourhood in 4×4 regions and compute orientation histograms along 8
directions
  SIFT descriptor: concatenation of all 4×4×8 (=128) values

Lec. 6
78
SIFT keypoints
 Final SIFT keypoints with detected orientation and scale

Lec. 6
79
Feature stability to view point change

Lec. 6
80
SIFT for Planar recognition
 Planar surfaces can be reliably recognized at a

rotation of 60° away from the camera
 Only 3 points are needed for recognition

 But objects need to possess enough texture

Lec. 6
81
Recognition under occlusion

Lec. 6
82
Place recognition

Lec. 6
83
Multiple panoramas from an unordered image set
SIFT is used in current consumer cameras (e.g. Sony, Canon) to build panoramas from multiple
shots!

Lec. 6
84
Demos
 SIFT feature detector Demo: for Matlab, Win, and Linux
(freeware)
http://www.cs.ubc.ca/~lowe/keypoints/
http://www.vlfeat.org/~vedaldi/code/sift.html
 Do your own panorama with AUTOSTITCH (freeware):

http://www.cs.ubc.ca/~mbrown/autostitch/autostitch.html

Lec. 6
85
Introduction to tomorrow‘s exercise
 How to Build a Range Finder using an Omnidirectional Camera

Lec. 6
86
Building a Range finder using an omni camera
1.  Acquire an image using the Matlab Image

Acquisition Toolbox.
2.  Calibrate the camera-mirror system:
detect the center of the image and
specify internal & external boundaries of
the omni image.
3.  Detect the wall contours using an
appropriate image processing technique.
4.  Calculate the distance to the walls using
the knowledge of the optics of the
camera-mirror system.
5.  Propagate the uncertainties from the
pixel points to the distances.

Lecture6 PDF

Uploaded by

Copyright:

Available Formats

Lecture6 PDF

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Lecture6 PDF

Uploaded by

Copyright:

Available Formats

Autonomous Mobile Robots

Zürich Autonomous Systems Lab

 Monochrome image  matrix of intensity values

 Images capture a lot of information

 Reduce the amount of input data:

© R. Siegwart , D. Scaramuzza and M.Chli, ETH Zurich - ASL

Image from http://www.flickr.com/photos/mukluk/241256203/

© R. Siegwart , D. Scaramuzza and M.Chli, ETH Zurich - ASL

Image from http://www.flickr.com/photos/mukluk/241256203/

© R. Siegwart , D. Scaramuzza and M.Chli, ETH Zurich - ASL

Image from http://www.flickr.com/photos/mukluk/241256203/

© R. Siegwart , D. Scaramuzza and M.Chli, ETH Zurich - ASL

 Edge / Corner Detection

© R. Siegwart , D. Scaramuzza and M.Chli, ETH Zurich - ASL

Lowpass filtered image Highpass filtered image

© R. Siegwart , D. Scaramuzza and M.Chli, ETH Zurich - ASL

Image Filtered Image

 For example, an averaging filter is:

© R. Siegwart , D. Scaramuzza and M.Chli, ETH Zurich - ASL

 Linear: every pixel is replaced by a linear combination of its neighbours

 Shift-invariant: the same operation is performed on every point on the image

A more detailed tutorial on this here: http://www.cs.umd.edu/~djacobs/CMSC426/Convolution.pdf

© R. Siegwart , D. Scaramuzza and M.Chli, ETH Zurich - ASL

 How to handle boundaries?

© R. Siegwart , D. Scaramuzza and M.Chli, ETH Zurich - ASL

© R. Siegwart , D. Scaramuzza and M.Chli, ETH Zurich - ASL

 Common practice for image smoothing: use a Gaussian

: controls the amount of smoothing

Normalize filter so that values always add up to 1

© R. Siegwart , D. Scaramuzza and M.Chli, ETH Zurich - ASL

 Common practice for image smoothing: use a Gaussian

99% : controls the amount of smoothing

Normalize filter so that values always add up to 1

© R. Siegwart , D. Scaramuzza and M.Chli, ETH Zurich - ASL

 Derivative of an image: quantifies how quickly intensities change

 Approximate a derivative operator:

© R. Siegwart , D. Scaramuzza and M.Chli, ETH Zurich - ASL

 Filter = template 3 8 3  test it against all image locations

 Similarity measure: Sum of Squared Differences (SSD)

Ignoring the filter values beyond the image-boundaries,

© R. Siegwart , D. Scaramuzza and M.Chli, ETH Zurich - ASL

 Filter = template 3 8 3  test it against all image locations

 Similarity measure: Sum of Squared Differences (SSD)

 Similarity measure: Correlation?

© R. Siegwart , D. Scaramuzza and M.Chli, ETH Zurich - ASL

 Filter = template 3 8 3  test it against all image locations

 Correlation value is affected by the magnitude of intensities

© R. Siegwart , D. Scaramuzza and M.Chli, ETH Zurich - ASL

 Filter = template 3 8 3  test it against all image locations

 For better invariance to intensity changes

© R. Siegwart , D. Scaramuzza and M.Chli, ETH Zurich - ASL

© R. Siegwart , D. Scaramuzza and M.Chli, ETH Zurich - ASL

 Example: Constant averaging filter

 If i.e. this is a square filter

 2D Correlation  no. multiplications per pixel =

This example was generated with a 21x21 mask

 Example: Constant averaging filter

 If i.e. this is a square filter

 2D Correlation  no. multiplications per pixel =

 Monochrome image  matrix of intensity values

 Images capture a lot of information

 Reduce the amount of input data:

 Edge / Corner Detection

  For example, an averaging filter is:

  Linear: every pixel is replaced by a linear combination of its neighbours

  Shift-invariant: the same operation is performed on every point on the image

 How to handle boundaries?

 Common practice for image smoothing: use a Gaussian

 Common practice for image smoothing: use a Gaussian

 Derivative of an image: quantifies how quickly intensities change

 Approximate a derivative operator:

 Filter = template 3 8 3  test it against all image locations

 Similarity measure: Sum of Squared Differences (SSD)

 Filter = template 3 8 3  test it against all image locations

 Similarity measure: Sum of Squared Differences (SSD)

 Similarity measure: Correlation?

 Filter = template 3 8 3  test it against all image locations

 Correlation value is affected by the magnitude of intensities

 Filter = template 3 8 3  test it against all image locations

 For better invariance to intensity changes

 Example: Constant averaging filter

 If i.e. this is a square filter

 2D Correlation  no. multiplications per pixel =

 Example: Constant averaging filter

 If i.e. this is a square filter

 2D Correlation  no. multiplications per pixel =

 Example: Constant averaging filter

 If i.e. this is a square filter

 2D Correlation  no. multiplications per pixel =

  We usually want to smooth by the same amount in both x and y directions

  So this simplifies to:

  Another separable filter

  Likewise, in 2D we flip the filter both horizontally & vertically

  Key difference between correlation and convolution is that convolution is associative:

  Ultimate goal of edge detection: an idealized line drawing.

  Edge contours in the image correspond to important scene contours.

  Edges correspond to sharp changes of intensity

  Change is measured by 1st order derivative in 1D

  Big intensity change  magnitude of derivative is large

  Or 2nd order derivative is zero.

  Image intensity shows an obvious change

  Where is the edge?  image noise cannot be ignored

  Where is the edge? Edges occur at maxima/minima of

  This saves us one operation:

  Discard pixels with (i.e. edge strength), below a certain below

  Non-maximal suppression: identify local maxima of

  We need to match (align) images

 Detect feature points in both images

 Detect feature points in both images

 Find corresponding pairs

 Detect feature points in both images

 Find corresponding pairs

 Use these pairs to align images

  Feature points are used also for:

  Robot navigation (we will see in tomorrow’s exercise)

  Object recognition (live demo)

  Image alignment (panoramas)

  Indexing and database retrieval -> Google Images or http://tineye.com

• Key property: in the region around a corner, image gradient has