Computer Vision and Pattern Recognition Technologies for Construction - Mobile and Pervasive Computing in Construction - page 195

Figure 10.4 Recognition of the personnel in construction sites (the white circle mark on the face is
to obscure individual identities) (Schiff et al., 2007)
priori knowledge about the shapes of vehicles, various kinds of vehicles, such as
sedan, hatchback, minivan, and so on, were modeled separately and recognized by
matching with the edge map of the images. Panin et al . (2008) also used 3D models
for recognizing a toy airplane. Their 3Dmodel was composed of color statistics and
intensity edges. Furthermore, they overcame the problem of high computational
costs by employing a GPU (Graphic Processing Unit)-accelerated contour sampler.
Georick et al . (1996) presented a feature-based model-free car recognition algo-
rithm. Instead of using a priori knowledge of shapes, they trained features of cars
with an artificial neural network. The experiment showed a near real-time process.
Sun et al . (2002) used the combination of wavelet and Gabor features to train the
characteristics of vehicles. Wavelet features provided edge information, which is a
good feature for vehicles, and Gabor features enabled it to be adapted to both
orientation and scale variation. Their experiment results showed accuracy enhance-
ment. Many researchers have applied principal component analysis (PCA) to
vehicle recognition (Sun et al ., 2004; Truong and Lee, 2009). PCA is one of the main
trends in feature extraction. One of the largest issues concerning PCA is how to
select the principal components. Sun et al . (2004) used genetic algorithms (GAs) for
selecting optimal components. This process could remove redundant features to
provide more efficient and effective recognition.
As mentioned, many types of equipment are involved in construction sites, and
independent detection processes are required for the detection of each type. This
problem can be addressed by combining recognition and segmentation. For
example, Shotton et al . (2008) proposed an algorithm known as the Semantic
Texton Forests (STFs) method. Basically, the method segments images and
categorizes each segment into pre-defined categories. The STFs method performs
based on the bag of semantic textons. It trains the characteristic of context
information as well as the objects' appearance. The STFs method is computation-
ally inexpensive, since it does not include complicated computations. Also,
segmenting and categorizing in pixel-wise comparison, it can identify various
types of entities in a single step. In Brilakis et al .'s (2010) vision tracking framework,
the STFs method was applied to construction equipment detection (Figure 10.5).
The method was trained to recognize separately the wheels and body parts of the
piece of equipment, thus allowing for the addition of contextual information
concerning the relative position between the two pieces of equipment. Their
method achieved about 80% pixel-wise accuracy.