The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part Bl. Beijing 2008
171
used in this paper. In the following sections, we briefly review
both methods.
2.3.1 RANSAC Method:
Dividing the images of both datasets into patches: Due to the
large size of two datasets which causes difficulties in running
the program large image data have been divided in smaller
patches.
The "RANdom SAmple Consensus" RANSAC was introduced
by Fischler and Bolles in 1981. If this concept is used in the
context of image registration the idea is to separate outliers and
inliers in a set of point pairs which are introduced to a
coordinate transformation. More generally speaking, the basic
assumption is that the data consist of inliers, i. e., data points
which can be explained by some set of model parameters, and
outliers which are data points that do not fit the model. In
addition, the data points are subject to noise. An advantage of
RANSAC is its ability to robustly estimate the model
parameters. It finds reasonable estimates of the parameters even
if a high percentage of outliers are present in the data set. A
small drawback of RANSAC is that a complete search would be
computationally very expensive. Therefore the number of
random samples which is selected to estimate the parameters is
usually limited by an upper number which may lead to a
suboptimal solution (Fischler and Bolles, 1981).
2.3.2 Baarda’s Data Snooping Method:
Baarda’s method is one of the most commonly used blunder
detection methods. It separates the inliers from the outliers
using the estimated normalised residuals. With observation
vector L and design matrix A the least squares estimate of the
unknowns x is found by
x = (A T A)~ l A T L (2-7)
The residual vector v and its covariance matrix C v
v = Ax- L
C,=crl{l-A{A T Ay x A T ) (2-8)
Designing a GUI: A GUI is developed which handles loading
datasets and specifying SIFT parameters, in particular: Number
of octaves, number of images in each octaves (levels), standard
deviation of Gaussian Function (a), threshold of SIFT matching,
and some others of minor importance. The error search with
RANSAC and data snooping is visualised to simply control the
impact of SIFT parameter settings onto the registration result of
Figure 1 : Flowchart of the registration program
are used to determine normalized residuals by computing the
ratio of a residual and the square root of corresponding diagonal
element of the covariance matrix C v
the four types of the LiDAR data used in the experiments (i.e.
LIDAR range, LIDAR intensity, both with first and second
pulse).
V
yj diag(Cç)ji
(2-9)
Normalized residuals follow the standard normal distribution
N(0,1) if no errors are present in the data. If a normalised
residual is above a critical value, the corresponding observation
might be erroneous. Data snooping is the process of eliminating
the observation, which produces the largest normalised residual.
This process has to be repeated until no further outlier is
detected anymore.
Overall registration procedure: Flowchart of the overall
registration process is shown in the following figure (Figure 1).
The histogram matching between LiDAR intensity and aerial
images serves for convenience of visual evaluation. SIFT
keypoints are extracted from the original image data. The
transformation parameters are used to rectify the target image to
the reference image frame.
3. EXPERIMENTAL INVESTIGATIONS
3.1 Test data set of Stuttgart city
The TopScan Company has acquired LiDAR data and aerial
images in April 2006 with a sampling density of 4.8 points per
square meter for the LiDAR data. Simultaneously aerial colour
images with 20 cm ground resolution have been recorded using
2.4 Procedural development
The procedural development undertaken in this research can be
summarized as follow: