529 
VEHICLE EXTRACTION USING HISTOGRAM AND GENETIC ALGORITHM BASED 
FUZZY IMAGE SEGMENTATION FROM HIGH RESOLUTION UAV AERIAL 
IMAGERY 
LI Yu 
Department of Geography, University of Waterloo, 602 White Cedar Ave., Waterloo, Ontario,Canada, N2V 2W2 - 
y621i@fes.uwaterloo.ca 
Commission III, WG III/5 
KEY WORDS: Vehicle Extraction, Histogram, Genetic Algorithm, Image Segmentation, Unmanned Aerial Vehicle (UAV) 
ABSTRACT: 
In this paper, an approach for extracting vehicles from UAV aerial imagery is given. The approach is based on a fuzzy segmentation 
algorithm, which combine fuzzy c-partition and genetic algorithm, and the geometric feature of vehicles for vehicle extraction. In 
this research, UAV colour imagery is examined experimentally. The results obtained demonstrate that most extracted vehicles match 
well with the original ones. From the analysis of results, it can be concluded that the proposed method is effective in both visual 
effect and positional accuracy. 
1. INTRODUCTION 
The effectiveness and usefulness of traffic management systems 
is dependent on accurately estimating the traffic flow conditions 
on traffic networks being monitored and managed. The ground 
based data are commonly available for such estimation. 
However, a limitation of the ground based methods for traffic 
data collection is that they rely on techniques that are strictly 
and spatially local in nature. For example, the cameras at fixed 
locations for traffic monitoring cannot easily observe the spatial 
progression and movement of traffic beyond the field of view 
since these camera platforms are not mobile. 
Airborne based data have the potential to significantly enhance 
the quality of traffic condition estimations due to their spatial 
scale and connectivity (Grejner-Brzezinska et al., 2007). To 
date there have been growing research activities in using of 
aerial imagery for transportation management, examples 
include the excellent summary in Kumar et al. (2001) and 
Angel et al. (2003). Most of researches examined the use of 
conventional aircrafts such as helicopter and manned light plane 
to collect aerial imagery. Unfortunately, they have a limit 
capability to acquire low cost and timely imagery. 
The use of high resolution aerial imagery acquired from 
unmanned aerial vehicle (UAV) would be attractive for 
studying, modelling, and monitoring traffic flow, since the 
UAV platform allow for the easy, cost effective and timely 
acquisition of high resolution and wide spatial coverage 
imagery unobtainable from ground based sensors. To gain 
acceptance for using UAV aerial imagery in traffic applications, 
it is necessary to demonstrate that vehicles can indeed be 
identified, extracted and classified accurately from the imagery 
and to establish a framework for data fusion with the ground 
based traffic information. 
In this light, the objective of this research is to develop vehicle 
extraction approach to demonstrate the ability to identify and 
extract vehicles from high resolution UAV imagery. This 
approach is following two stages for vehicle extraction. Firstly, 
the colour UVA aerial imagery is segmented by a proposed 
segmentation algorithm. The designed segmentation algorithm 
is based on fuzzy c-partition, in which the colour histogram is 
used for colour space analysis to obtain a proper initial estimate 
of centre positions, then genetic algorithm is employed to 
optimally cluster the colour space data point projected from an 
input colour imagery, hence optimal colour segmentation can 
be achieved. Secondly, the post-processing procedure is carried 
out on segmented image by use binary mathematical 
morphology operators to extract the outlines of vehicles. 
The paper is organized as follows. In the next section, the basic 
concepts from fuzzy c-partition colour histogram, genetic 
algorithm and colour histogram are introduced. In the third 
section, the proposed approach to vehicle extraction is 
described. The approach is applied to UAV aerial colour 
imagery in the fourth section. Conclusion about the approach is 
given in the fifth section. 
2. BACKGROUND 
2.1 Fuzzy C-Partition 
According to the fuzzy paradigm (George and Bo, 1995), fuzzy 
clustering is seen as partitioning a data space into a number of 
fuzzy sets and assigning each data point a membership to each 
cluster. Consider a vector set V formed by n vectors in L 
dimensional real number space R L , i.e., V = {Vi, ... , v„}, vy = 
(y,i, ..., Vj L ) E R l and j — 1,2, ...,«, a fuzzy c-partition on V is 
represented by the fuzzy partition matrix P = \py\, i = 1, ..., c, 
where c is the positive integer to indicate the number of the 
clusters in the partition, and p,y e [0, 1] is the fuzzy membership 
value of v 7 belonging to zth cluster and satisfy, 
= 1, for ally =1,...,« 
i=l 
(1) 
0 < Yp, < n, for all z = l,..,c 
(2) 
7=1 
Before using fuzzy c-partition to design a clustering algorithm, 
the following two issues should be solved. First one is how to