The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B5. Beijing 2008 
585 
4. PROPOSED ALGORITHM 
Nowadays, almost all the terrestrial laser scanners can 
return the distance from the point on surface and the energy of 
the backscattered laser light in this point for each measurement. 
This leads to two different sets of data. The 3D data are 
recorded from the distance measurement, whereas a panorama 
image can be generated from the reflectivity information. We 
will refer to this image as the reflectance intensity image for it 
looks similarly to a real intensity image taken by cameras. 
4.1 Preprocessing of Reflectance Intensity Images 
The reflectance intensity image is generated from the 
backscattered laser light which is a signal of high dynamic 
range. The strength of the return varies over a large range, from 
almost no return due to low reflective, far away surfaces, to 
direct reflection from retro reflective material. For the Riegl 
LMSZ360I this fact is accounted for by storing the reflectance 
information as 16-bit numbers. Since most of the displays and 
many standard image processing tools are still designed for 8- 
bit image data, we decide to convert the reflectance information 
to 8-bit. As shown in Figure 1, due to lost information in the 
course of conversion from 16-bit data, the 8-bit reflectance 
intensity image has characteristic of low contrast (as shown in 
Figure 1(a)). Therefore, we have to firstly apply image 
preprocessing to make this low contrast image appear more like 
a typical intensity image. 
Histogram equalization and normalization are usual tools for 
increasing the contrast of images, especially when the usable 
data of the image is represented by close contrast values. 
Histogram equalization and normalization can be outlined as 
follows: 
1. Histogram equalization accomplishes increasing the contrast 
of images by effectively spreading out the most frequent 
intensity values. A disadvantage of the method is that it is 
indiscriminate. It may increase the contrast of background noise, 
while decreasing the usable signal. Consider a reflectance 
intensity image, and let n be the number of occurrences of the 
gray level i. The probability of an occurrence of a pixel of 
level i in the image is 
/?(/) = «./«,/6 0,...» L-1 
(2) 
L being the total number of gray levels in the image, n being the 
total number of pixels in the image, and p being in fact the 
image’s histogram, normalized to [0, 1]. Let us also define c as 
the cumulative distribution function corresponding to p, defined 
by: 
f 
c(0 = £p( x i) ( 3 ) 
j=0 
where c also known as the image’s accumulated normalized 
histogram. We would like to create a transformation of the form 
y= T(x) that will produce a level y for each level x in the 
original image, such that the cumulative probability function of 
y will be linearized across the value range. The transformation 
is obtained by: y, = T(xj = c(i). Notice that the T maps the 
levels into the domain of [0, 1]. In order to map the values back 
into their original domain, the following simple transformation 
needs to be applied on the result: 
y. = y {Max-Min) + Min 
(4) 
2. Histogram normalization stretches an image’s pixel values to 
cover the entire pixel value range (0 - 255). The intensity image 
is preprocessed by subtracting the minimum grey value from 
each pixel and dividing by its max-min range. Visually the 
image appears to have increased in contrast. 
y. = ((y_ - Min) /{Max - Min)) x 255.0 
(5) 
Böhm and Becker (Böhm and Becker, 2007) apply histogram 
equalization to increase the contrast of reflectance image. Then 
they extract the SIFT features and match these features in 
equalized reflectance image. However, in this paper, we prefer 
applying histogram normalization instead of equalization 
because the histogram normalization operator does not increase 
the contrast of background noise which usually leads to false 
matches. After this preprocessing, we take the advantage that 
we can rely on a standard implementation for feature extraction 
and do not have to alter lots of parameters. 
4.2 SIFT Feature Based Key Points Matching 
The Scale Invariant Feature Transform (SIFT) developed by 
Lowe (2003) is invariant to image scale and rotation, and 
provides robust matching across a substantial range of affine 
distortion, change in 3D viewpoint, addition of noise, and 
change in illumination. Our application employs a standard 
SIFT feature extraction and key point matching based on those 
features. For example, Figure 1(c) shows 301 matches obtained 
from 12093 extracted SIFT feature points. 
4.3 Geometric Constraint 
Due to invalid points, holes, dark or reflective spots on the 
object’s surface, especially symmetry and self-similarity of the 
façade structures in the scans, the pairs of matched points 
contain a lot of false matches. As shown in Figure 1(c), 
repetitive elements such as windows and bricks on the ground, 
which are especially dominant in the example scene, cause false 
matches, when the geometry of the scene is ignored. 
Original reflectance intensity image 
Matching result from equalized reflectance image