THE AUTOMATIC RECOGNITION, LOCATION AND LABELLING OF TARGETS IN DIGITAL 
PHOTOGRAMMETRIC ENGINEERING MEASUREMENT 
J. Chen & T.A. Clarke 
Department of Electrical, Electronic, & Information Engineering, 
City University, Northampton Square, 
London, EC1V OHB, 
ENGLAND. 
ABSTRACT: 
A major problem in digital photogrammetric nmalysis of engineering structures using multiple views and bundle 
adjustment techniques 1s the precise recognition, 
ocation, and labelling of targets. This task is complicated by: 
variable surface reflectivity, non-ideal illumination, occlusion, and the variation of target size with distance from 
camera stations. In this paper the authors discuss the use of mine images of a targeted section of a wind power 
generator blade to automatically calculate the coordinates of t 
e targets. Three aspects are discussed: target 
recognition using a binary image, subpixel location using a grey scale image, and unique labelling. 
(i) The target recognition process uses a binary matching technique where prior knowledge of the target 
characteristics are used to reject spurious targets. 
(ii) The local region around each target image is independently analysed and an adaptive location algorithm 
used to give a subpixel estimation of the position of the target. 
(ii) A small number of control targets are used to perform a transformation of the camera images to enable 
reliable automatic labelling of each unique target image. 
Results from applying these methods to provide the coordinates of a targeted section of a wind power turbine blades 
are presented, and the accuracy and efficiency analysed. 
KEY WORDS: Targets, Recognition, Image-processing, Subpixel, 3-D. 
1. INTRODUCTION 
In digital photogrammetry, using multiple camera 
views, the 3-D spatial relationship of targets which are 
placed on an object can be determined using the bundle 
adjustment technique from the 2-D spatial locations of 
the targets in each camera view. An important 
consideration, if the full potential of this system is to be 
exploted. is unique Hed identification which must 
take place reliably if the resulting 3-D spatial 
coordinates of the targets are to have any meaning. It is 
common for this process to take place by moving a 
cursor to the target and labelling it according to some 
predetermined scheme. However, this method requires 
a skilled operator and is time consuming with the 
possibility of poor reliability. Another aspect is the 
precision with which the 2-D coordinates of the targets 
can be obtained. 
With the development of computer vision and pattern 
recognition techniques the use of digital image 
processing of CCD camera produced data is being more 
widely used (Gruen, 1989). Bethel, 1990, discusses the 
fact that photogrammetrists might benefit by looking 
towards computer vision techniques where some 
methods have become established over many years for 
dealing with digital imagery (Bhanu, 1986). The 
methods applied here rely m on a machine vision 
approach to the problems of target recognition such as 
described by West, 1988. The machine vision 
community primarily works in image space to perform 
task such as: matching, pattern recognition, target 
location, edge detection, etc. Photogrammetrists are 
more used to mapping image coordinate information 
back into the object space. Some cross fertilisation may 
be of mutual benefit. This paper represents the work of 
two machine vision researchers who are attempting to 
learn from the photogrammetric community. 
This paper deals first with target recognition using a 
binary image and target location using grey scale 
images. The aim is the detection of the coordinates of 
the target object with respect to the sensor. The second 
area sr is considered is the consistent labelling of 
the targets with respect to a reference image and 
subsequent images which contain views of the same 
target. The method proposed uses the best features of 
each method for an optimal solution. Research into 
binary and grey scale target location have been 
performed elsewhere (Mikhail, 1984; Rosenfield, 
1988). However, in this paper both are considered along 
with the labelling problems for multiple views. 
A novel feature of this investigation is the use of the 
Bundle Adjustment technique as a means of comparing 
the accuracy of target location methods. The aim of the 
investigation is the automatic 3-D measurement of a 
section of a wind turbine blade. The research results 
indicate that these methods can efficiently: recognise 
the targets, locate them to subpixel accuracy, and 
properly label the them in the multiple images to 
produce a 3-D range map of these targets. 
The study uses a PC-AT computer, a Pcvision 
Framestore Board, and some general purpose image 
processing software. The image size and resolution is 
512x512x8 Bits. All the software is written in the C 
language except the interface software, which is written 
in Assembly language.