Civil Engineering Reference
In-Depth Information
et al
., 2004). The second phase is to find whether the extracted features contain the
pre-defined object's feature through the use of best-bin-first search, multiple
randomized k-d trees or a hierarch k-means tree (Nist
enius, 2006;
Silpa-Anan and Hartley, 2008). The methods are powerful in detecting a specific
object, but they are not appropriate for object category detection (Zhu and
Brilakis, 2010a). For example, these methods can match a specific column with
itself in another image, but cannot effectively detect similar ones. Structural
elements, although geometrically simple, are characterized by large topographical
variations (e.g. aspect ratio), and therefore no simple scale/affine transformation
can characterize them.
er and Stew
10.2 Construction equipment and personnel recognition
In large scale construction sites, there are usually a large number of project-related
entities such as workers and equipment. The congested environment of construc-
tion sites makes it difficult to recognize these entities, which is sometimes required
for construction management. For example, it may be demanding to calculate the
number of workers or equipment involved in a certain construction activity or the
distance between workers and equipment to prevent their collision. In practice,
Gong and Caldas (2010) employed a vision-based object recognition algorithm for
the automated productivity analysis. Also, the vision tracking methodology pre-
sented by Brilakis
et al
. (2010) requires automatic object recognition as an
initialization of their framework. Accordingly, the recognition of project-related
entities can help automate construction management tasks.
Recognizing workers in images is possible using human recognition methods, on
which vigorous research work has been, and continues to be, performed in the area
of computer vision. Human recognition is one of the more difficult topics within
the subject of object recognition. What makes it more difficult than others is the
variety of human appearances. Compared to rigid objects such as vehicles, people
can be imaged in extremely different ways depending on their poses, the colors of
garments, and so on. This variety of appearances requires object models to have a
higher level of adaptability, which is hard to accomplish with a limited number of
template models. For this reason, a machine learning process is generally intro-
duced to assist in human recognition. By training of human features, a classifier is
established that can determine whether or not the features of a given image area
belongs to a human. (Image features are extracted from a large number of training
images and are used to construct a classifier through machine learning algorithms,
such as SVM, K-nearest neighbors, and artificial neural network, etc. This process is
referred as “training”. For example, various vehicle images (positive images) and
non-vehicle images (negative images) can be loaded into a machine learning
algorithm to construct a classifier which can recognize a vehicle in another image.)
Dalal and Triggs (2005) made a remarkable advance in human recognition,
which outperformed other previous works in terms of accuracy. They proposed
histogram of oriented gradient (HOG) as a novel description of human features,
which turned out to be powerful with respect to discerning human features. Their
Search WWH ::
Custom Search