Você está na página 1de 11

Introduction to Sensor Interpretation

What is sensor interpretation? - abstraction of high-dimensional input spaces to low-dimensional output spaces Why interpret sensors? - working with large datasets is expensive - smoothing and filtering based on models can yield better results in applications

From Ziegler et al., 3D Reconstruction Using Labeled Image Regions


3/19/2013 CS225B Kurt Konolige

Some Issues
Models

A model is a mathematical description of the abstract feature, controlled by a few parameters

Sensor Noise

Noise is the deviation of a sensor reading from its ideal value Noise complicates the process of feature extraction, and makes probabilistic techniques useful

Graphs from Guo, A Multiple-Line Fitting Algorithm without Initialization


3/19/2013 CS225B Kurt Konolige

Some Issues
Segmentation

What is the important part or parts of a set of sensor readings? Figure/ground distinction [Edgar Rubin]

Outliers / Data association

Some readings are not associated with the feature at all, and do not correspond to normal noise values

[Roger Shepard] Graph from Guo, A Multiple-Line Fitting Algorithm without Initialization

3/19/2013

CS225B Kurt Konolige

Techniques: Maximum Likelihood


Maximum Likelihood Method [Linear Regression]

Measurement function f i ( x) zi i Independent measurements zi with value z Gaussian distribution of error, with the same variance for each zi THEN
maxx p(x|z) is given by

i ) 2 min ( f i ( x) z
x

Line parameters: slope, intercept s,i

yx

Measurement equation:

f x (s, i)

f x (s, i) y

3/19/2013

CS225B Kurt Konolige

Techniques: Maximum Likelihood


Maximum Likelihood Issues

Which model? - some are better than others

Outliers Gaussian assumption means outliers pull a lot of weight


Solutions are in robust statistics

i ) 2 min ( f i ( x) z
x

Ignore high residuals Least Median of Squares [Rousseeuw] Consensus estimators

min median ( f i ( x) zi ) 2
x

s,i

Unequal Variance

Some measurements should count more more than others Mahalanobis distance:

yx

f x (s, i)

( f i ( x) zi )T C 1 ( f i ( x) zi )

3/19/2013

CS225B Kurt Konolige

Techniques: Incremental/Recursive Line Fitting


Recursive Line Fitting

Order points along the curve Start with a straight line connecting the first and last points Find point with max distance Split and apply algorithm to each segment

Incremental Line Fitting

Order points along the curve Add next few points to a new line Fit line and check residual If residual is good, add next point and re-fit If residual is bad, declare line done and start a new line

3/19/2013

CS225B Kurt Konolige

Techniques: Incremental Line Fitting

From T. Darrell course notes

3/19/2013

CS225B Kurt Konolige

Techniques: Hough Transform


Slope/intercept parameterization of a line
sin(a) x cos(a) y d 0

A point in (a,d) represents a line in (x,y) For a give point (x,y), there is a family of lines going through the point a curve in (a,d)

From T. Darrell course notes

3/19/2013

CS225B Kurt Konolige

Techniques: Hough Transform


Issues

sin(a) x cos(a) y d 0

Representing curves in (a,d) typically use a grid of cells For each point in (x,y), increment the count in corresponding cells Find loci Noise, resolution, and computation are problems

From T. Darrell course notes


3/19/2013 CS225B Kurt Konolige

Techniques: RANSAC
RANdom SAmple Consensus

[Fischler and Bolles]

Uses consensus to determine data association Uses random sampling to find seeds for consensus Robust to outliers Can find multiple features
How many pairs must we pick before we find a good line?

3/19/2013

CS225B Kurt Konolige

Techniques: RANSAC
Algorithm

[Fischler and Bolles]

Function f(x) = y for finding a model y from a set of data points x of size n (e.g., in the case of lines, two or more points) Choose a random set x Find all inliers to model y Accept if inliers > threshold (maybe replace previous model, or add to models if using multiple models) Iterate for some number k

How many pairs must we pick before we find a good line?

z (1 w n ) k
k log( z ) log(1 w )
n

z is the probability of only bad picks

k is chosen based on z

3/19/2013

CS225B Kurt Konolige

Você também pode gostar