the exposure time of each image is very short due to the high
frame rate measurement, the consistent and homogeneous il-
lumination condition is of great importance to subsequent im-
age processing. High-power lamps can be placed close to the
cameras if the natural light is insufficient. On the other hand,
as textural features are not salient in the case of complex ex-
perimental conditions (e.g., the major model material is quartz
sand in the following experiments), natural targets detected
by interest operator are sparse or hard to obtain. Therefore,
artificial targets are commonly applied in videogrammetric
measurement for the sake of efficiently monitoring the key
positions of the structural models. In this study, flat circular
targets are used as control points and tracking points. Their
shape information can provide accurate spatial point coordi-
nates and is beneficial to the image sequence processing. With
regard to the control targets, cross-wire, and reflective sheet
are added to the center in order to improve the precision of the
total station measurement. The employed targets can be made
of retroreflective material or printed on other inexpensive
materials. However, the retroreflective targets require the right
lighting conditions and are ineffective in the case of weak lu-
minance as well as large entrance angle (Burgess
et al.
, 2011).
The core components of the videogrammetric hardware
system are high-speed cameras and a computer. For video-
grammetric measurement, cameras play the most important
role, and their capabilities directly decide the potential. A
higher accuracy of dynamic responses can be guaranteed
when cameras with a higher resolution and higher frame rate
are used. In general, the selection of the appropriate cameras
is driven by the application requirements considering cost,
resolution, acquisition speed and frame rate, synchronization,
the amount of data, field of view, image scale, etc. (Luhmann,
2010). The function of the computer in the measurement sys-
tem is to adjust, synchronize, and control the cameras as well
as store and process the image data. A high-performance data
acquisition card and a synchronous controller are integrated
into the computer to facilitate the fulfillment of these tasks.
Real-time image sequence data acquisition is ensured by
means of an efficient videogrammetric hardware system.
Image Sequence Processing
This section introduces the proposed videogrammetric image
sequence processing, including image preprocessing, target
recognition and matching, target tracking, as well as 3D spatial
coordinate calculation. The objective is to calculate 3D coordi-
nates of the key positions of the structural models at each epoch.
Image Preprocessing
As a subpixel accuracy is pursued, good illumination and
contrast conditions should be strictly demanded. Therefore,
image enhancement is performed on every frame of the ac-
quired image sequence data. The Wallis filter, which adap-
tively and locally adjusts the pixel intensity value to match
the target mean and standard deviation, is employed to enrich
the texture details and increase the contrast levels of images
(Barazzetti and Scaioni, 2010).
Target Recognition and Matching
The next step is target recognition and matching in the first
frames of the stereo image sequences. We first detect the
targets, including tracking points and control points, using
ellipse detection based on
a contrario
theory and least squares
ellipse fitting, followed by determining their correspondences
in the two images using point set registration.
For the purpose of accurately determining the central posi-
tions of targets, a coarse-to-fine detection scheme is adopted
in this study. In the stage of coarse detection, a state-of-the-
art ellipse detector denoted as
ELSDc
(P
ă
tr
ă
ucean
et al.
, 2017)
is adopted because of its convenience and robustness in
the presence of noise, cluttered backgrounds, and varying
illumination. There are three essential stages, i.e., candidate
generation by region growing and region chaining, candidate
validation, and model selection using an
a contrario
frame-
work. For a detailed description on the implementation and
superiority of
ELSDc
the reader is referred to P
ă
tr
ă
ucean
et al.
(2017). As several useless elliptical primitives exist in the out-
put, we filter the results by restricting the area and delimiting
angle of the outputs within the given possible range to obtain
the rough positions of the targets.
However, these positions are not sufficiently precise, and
thus, a local refinement considering the elliptical geometric at-
tributes is further needed to improve the target positions and to
remove the false detections. A target location method similar
to the method in Liu et al. (2015) is adopted in the stage of fine
detection. For each coarse estimate, an image block centered at
the rough position from the coarse detection is first extracted
with the size of double the estimated diameter. Gaussian low-
pass filtering is used to suppress the noise, and a Sobel edge
detection operator is used to detect the contours. Afterwards,
the image morphology algorithm (i.e., closing) is employed
and the holes within the image block are filled. The elliptical
targets are recognized according to the geometric attributes,
including area, perimeter, and degree of circularity. Only con-
nected regions whose geometric attributes locate within the
predefined ranges are identified as elliptical targets. Finally,
the central pixel coordinates of the identified ellipses are cal-
culated using a least squares fitting method (Matsuoka, 2014).
After separately estimating the target positions from
the stereo images, their correspondence in the stereo im-
ages should be subsequently established. The use of special
coded targets can facilitate automated target correspondence
determination. However, the coded targets are unsuitable for
scenes with large range and medium- or low-resolution sen-
sors, as the required coded targets should be inconveniently
large (Luhmann
et al.
, 2016). The suitable image processing
algorithm is therefore crucial in these instances. Normally,
this task is realized by a feature descriptor and feature match-
ing using neighboring intensity information. However, this
traditional strategy is inapplicable in our case due to the same
target pattern and the textureless model material as well as
Figure 2. Framework of the videogrammetric system.
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
September 2018
583
