11.07.2014 Views

A Probabilistic Approach to Geometric Hashing using Line Features

A Probabilistic Approach to Geometric Hashing using Line Features

A Probabilistic Approach to Geometric Hashing using Line Features

SHOW MORE
SHOW LESS

Create successful ePaper yourself

Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.

CHAPTER 2. PRIOR AND RELATED WORK 21<br />

is, the measured values are assumed <strong>to</strong> be distributed according <strong>to</strong> a Gaussian distribution<br />

centered at the true values and having standard deviation ç. More precisely, let èx i ;y i èbe<br />

the ëtrue" location of the i-th feature point in the scene. Let also èX i ;Y i è be the continuous<br />

random variables denoting the coordinates of the i-th feature. The joint probability density<br />

function of X i and Y i is then given by:<br />

fèX i ;Y i è= 1<br />

2çç 2 expè,èX i , x i è 2 +èY i , y i è 2<br />

2ç 2 è:<br />

Using this error model, he formulates a weighted voting scheme for the evidence accumulation<br />

in geometric hashing.<br />

He uses a weighted contribution <strong>to</strong> the modelèbasis<br />

hypothesis of an entry ç in the hash space based on a scene point's hash ç in the same<br />

space, <strong>using</strong> a formula of the form<br />

logë1 , c + A expè, 1 2 èç , çèæ,1 èç , çè t èë;<br />

where c and A are constants that depend on the scene density and æ is the covariance<br />

matrix in hash space coordinates of the expected distribution of ç based on the error<br />

model for èx; yè, the feature point in the scene.<br />

He shows that when weighted voting is used according <strong>to</strong> the above formula, the evidence<br />

accumulation can be interpreted as a Bayesian aposteriori classiæcation scheme. He<br />

shows that the accumulations are related <strong>to</strong><br />

logëProbèH k jE 1 ; æææ;E n èë<br />

where the hypothesis H k represents the proposition that a certain modelèbasis occurs in<br />

the scene and matches the chosen basis, and E i 's are pieces of evidence given by scene<br />

invariants.<br />

The above formulation is implemented on a 8K-processor Connection Machine èCM-<br />

2è and can recognize objects that have undergone a similarity transformation, from a<br />

library of 32 models. The models used for experiments are military aircraft and production<br />

au<strong>to</strong>mobiles. <strong>Features</strong> are extracted by <strong>using</strong> the Boie-Cox edge detec<strong>to</strong>r ë9ë and locating<br />

the points of high curvature. The features are the coordinate pairs of these points ë44ë.<br />

In this thesis, we also make use of a Bayesian aposteriori maximum likelihood recognition<br />

scheme based on geometric hashing. We use somewhat simpliæed versions of the<br />

formulae given by Rigoutsos. We make use of line features as opposed <strong>to</strong> point features.

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!