FEATURE EXTRACTION FROM MOBILE MAPPING IMAGERY SEQUENCES USING GEOMETRIC CONSTRAINTS 
Fei Ma and Ron Li 
Department of Civil and Environmental Engineering and Geodetic Science 
The Ohio State University, USA 
ma.85@osu.edu, li.282@osu.edu, 
KEY WORDS: feature extraction, edge detection, watershed transformation, GPS/INS, perspective geometry 
ABSTRACT 
This paper presents research results on feature extraction from mobile mapping imagery sequences using geometric constraints. A 
comprehensive feature extraction method is developed, where first, edges are detected using a multi-scale edge detector that combines 
first-order and second-order derivatives. Then multi-level thresholds are calculated to segment images. The result is used for further 
watershed transformation applied to improve the segmentation result. After edge selection and line linking, for example, preliminary 
road lines are detected with the help of GPS/INS data. In addition, based on perspective geometry, vertical and directional horizontal 
line segments are detected and used as seeds for extracting other objects. 
1. INTRODUCTION 
D ai = F(x, y) * T a (x, y) ,G 0i 
(4) 
Feature extraction is an area of active research in both 
photogrammetry and computer vision. In mobile mapping, it is 
also a critical step of object recognition. In the past, feature 
extraction has been divided into several stages, namely, low level 
image processing, edge detection, contour derivation, and shape 
modeling. The classical method is weak in dealing with complex 
scenes such as mobile mapping imagery. Later models, such as 
the deformable contour model (Kass et al. 1987), treated these 
problems in a general unified manner. They were used for road 
extraction from land-based and aerial images (Gruen and Li 
1997, Tao et al, 1998). This paper presents recent results of our 
study on feature extraction from mobile mapping imagery 
sequences through a multi-level approach using geometric 
constraints derived from GPS/INS data. 
2. A COMPOUND EDGE DETECTOR 
In our study, we combine LoG (Laplacian of Gaussian) and Drog 
(Derivative of Gaussian) operators into a compound edge 
detector (Li et al. 1998) to take advantages of information from 
second and first order derivatives. Mathematically, zero crossings 
of an image F(x,y) with two scale parameters are: 
G-. =F(x,y)*H^x,y) 0) 
G ai =F(x,y)*H 0i (x,y) (2) 
where * denotes convolution operation, H(x,y) represents LoG 
function, and oj and <?, are scale parameters (o' 1 < o\). 
Additionally, Drog operations with the constraint of G a are 
where T^y) represents the derivative of Gaussian. 
Thus, the edge detection works at two different scales. One for 
extracting the raw shape of an object, called initial resolution and 
the other for extracting the exact shape of the object and for 
distinguishing it from other similar objects. This is implemented 
by defining the initial scale <j, based on the result of the optimal 
edge detector of Canny (1986). The refined scale cr 2 is chosen 
according to the desired details. Edges detected at the initial 
resolution are intensified by edges detected at the refined 
resolution. Spurious noise is avoided by intersecting the results 
of the refined scale with that of the initial scale. 
Figures 1(a) to (d) are two stereo pairs of mobile mapping 
imagery from a sequence. Each image has 720 x 400 pixels. The 
pixel size is 0.0116 mm x 0.0136 mm. Left camera focal length is 
6.1288 mm and right camera focal length 6.1278mm. 
(Pair I) 
(a) Left 
I) 
(c) Left image (Pair II) 
(d) Right image (Pair II) 
Figure 1. Two pairs of mobile mapping images 
D a< =F(x,y)*T a Uy)-G a 
(3) Figures 2(a) to (d) are edge detection results at initial scale on 
four images in Figure 1. 
7A-2-1