Machine Vision for Robots Modeling Cameras Processing Images – Geometric/Moment based Features – Other Features.

Machine Vision for Robots Modeling Cameras• Processing Images

– Geometric/Moment based Features– Other Features

Histograms and Thresholding

Histogram: graphical presentation of the frequency count of the occurrence of each intensity in an image

Magnitude of histogram at a specific pixel value, hist(g), is the probability of the gray value, g, occurring in any picture element in the frame


Histogram: determine if image has sufficient contrast equalize image histograms to perform operations

(e.g. subtraction) determine appropriate threshold values


Threshold: Segmentation of the image (classify each pixel)

according to gray value – I.e. for threshold, T

Puts pixel into the set of background Sb or foreground Sf pixels

binary segmentation of the image

Connect Component Analysis Identify individual components (blobs) within

the image: Input: a binary image with objects having value 1

and background having value 0. Output: an image with each region having value

equal to the region label (and the background having value 0).

• 4-connected vs. 8-connected• A is not connected to C in 4-connectivity• A is connected to C in 8-connectivity

A

C D

B

Connect Component Algorithm One Algorithm (4-connectivity):

Pass 1: If A==1 and D==1 then EqualLabels (A,D) If A==1 and B==1 then EqualLabels(A,B) If A==1 and Both(B==1) and (D==1) then If (Label(B) != Label(D) then EqualLabels(B,D)

Pass 2: Relabel each pixel within each equivalence class

A

C D

BIn Matlab™ the routine bwlabel performs a connected component analysis

Geometric Properties Assume: binary image with ONE object having

value 1 and background having value 0.

Moment definition:

Geometric Properties Assume binary image with ONE object having

value 1 and background having value 0.

Moment definition:

Note m00 is the area

Geometric Properties Center of mass: point such that if all the

object’s mass were concentrated at that point, the first moments would not change

Geometric Properties Center of mass: point such that if all the

object’s mass were concentrated at that point, the first moments would not change

Note m00 is the area

The zero and first moments give the centroid

Geometric Properties Centralized Moment definition:

Note m00 is still the area but the first centralized moments are now zero

Geometric Properties Orientation: axis that passes through the object

such that the second moment of the object about that axis is minimal Or where

= minimum distance from pixel (c,r) to the line

To solve? Parameterize line with (, ) Compute partial derivatives of w.r.t. (, ) Find minima by setting partial derivatives to zero

Geometric Properties Line parameterization

Normal to line is Perpendicular distance from line to origin is

Now our task is to find:

Geometric Properties Derivative

Set to zero

Defines a line through centroid shift origin Line is now

Geometric Properties Derivative, set to zero

Solve (algebra…)

Or, using double angle formulas

Geometric Properties Principal axes from second moments

Orientation of principal axis w.r.t. image x-axis

Or using double angle formula we can also write:

Geometric Properties Alternate point of view

Think ellipses again Eigenvectors along major/minor axes Eigenvalues give length of major/minor axes

Geometric Properties Area, Centroid

Principal axes from second moments (centralized) Orientation

Major/ Minor axis length

Geometric Properties Normalized moments

Moment Invariants Functions whose values are approximately

constant under various transformations E.g. the zeroth moment does not change under

rotation about the optical axis (i.e. rotation in the image plane)

Geometric Properties Moment Invariants

Image Features

Measured quantities – can be used to “represent” the object Centroid Best fit ellipse (centroid + axes) Area Compactness

Perimeter2/(4 Area) ¸ 1 Eccentricity

(major axis length)/(minor axis length) Moment Invariants # Holes

We can represent the features associated with an object as a vector

Features & Patterns

area compactness Mi4 Mi5

We can represent the features associated with an object as a vector

A set of samples, described by their statistical properties with respect to a number of features defines a pattern.

Features & Patterns

From a set of samples we can find the mean value of each feature

The covariance is

Features & Patterns

The covariance is

Features & Patterns

The statistics (mean & covariance) can be used to classify an object.

Given a set of samples for objects (O1,…,OM) we can compute for each object class

For a new object, onew, we measure the features

Features & Patterns

Pattern classification: Which object (O1,…,OM) is this new object closest

too? How do we measure distance?

Euclidean:

Features & Patterns

Two features measured for two objects

Example: Euclidean Distance

Feature 1

Fea

ture

2

O1 O2

Compute statistics


Feature 1

Fea

ture

2

O1 O2

Compute statistics


Feature 1

Fea

ture

2

O1 O2

Measure features for new object


Feature 1

Fea

ture

2

O1 O2

Measured feature values for new object

Compute distance


Feature 1

Fea

ture

2

O1 O2

Euclidean distance to mean of features for Objects 1 and 2

Statistically – which object is more likely?


Feature 1

Fea

ture

2

O1 O2

Pattern classification: How do we measure distance taking into account

statistics? Inversely weight distance with variance Mahalanobis distance:

Features & Patterns

Weights the distance by the inverse of the (co)variance

Two features measured for two objects

Example: Mahalanobis Distance

Feature 1

Fea

ture

2

O1 O2

Compute statistics

Example: Mahalanobis Distance

Feature 1

Fea

ture

2

O1 O2

Use statistics to measure disanceExample: Mahalanobis Distance

Feature 1

Fea

ture

2

O1 O2

For M objects we need to store the mean of the feature vector and the covariance

For a new object, onew, we measure the features and compute the distance to each object class (M distance computations) and find the minimum

Features & Patterns

Machine Vision for Robots Modeling Cameras Processing Images – Geometric/Moment based Features – Other Features.

Documents