The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B5. Beijing 2008
3.2 Variable depth and width of the mortar channel
Typically the mortar between bricks is indented from the
surface of the bricks. This indentation is called the mortar
channel. When the mortar channel is deep, brick surfaces easily
stand out from the surface of the wall. In a laser scan this
appears as a surface discontinuity between the points on the
mortar and the points on bricks. If the mortar channel is shallow,
there is no surface discontinuity between mortar and brick, and
the segmentation yields segments that contain both mortar and
brick, i.e., under segmentation. In masonry walls, very often
bricks are packed tightly against each other and the gaps are
filled with mortar. Because of this, segmentation can yield
segments containing more than one brick.
3.3 Similar surface covering
Ideally, if bricks and mortar are made of different materials, it
should be possible to segment a wall based on the rgb values
(colour) of points. In practice this is not possible if bricks and
mortar are covered by paint, lichen, moss or a combination.
This gives bricks and mortar the same appearance or worse a
patchy appearance. Another variation on this problem is when
the bricks and mortar are made of the same material. For
example mud walls.
The above problem might be overcome by acquiring multi-
spectral images of a wall.
3.4 Discontinuities in brick surfaces
Segmentation assumes that a surface discontinuity only exists
between the surface of mortar and bricks. Bricks in a masonry
wall sometimes have surface discontinuities within them. This
leads to an over segmentation which may then complicate the
detection of whole bricks.
3.5 Surface lighting
Depending on the lighting parts of bricks and mortar may be in
shadow or highlighted. This complicates the segmentation of a
wall based on the shading/brightness of objects. Furthermore, it
complicates segmentation based on rgb, since brightness is
embedded in the rgb components. Converting rgb values to hsv
(hue, saturation and value) should help to untangle the effect of
the variability of surface lighting from the hue.
Preferably a wall should be scanned under diffused illumination
as available on a cloudy day. However, this may not always be
possible.
3.6 Variable resolution of point cloud
Choosing the parameters of segmentation assumes that the local
geometric surface characteristics of brick and mortar surfaces is
uniform throughout the point cloud. In practice this is not the
case for the following reasons:
- The edge of a scan will have a lower resolution than the
centre of the scan,
- The overlap area of two or more scans will have a higher
resolution than other areas of the scan.
- Objects at different distances from the scanner will have
different scan resolutions, as shown in
Figure 1. This applies to walls that are embedded, i.e.,
walls at different distances from the scanner.
Furthest wall is of a
lower resolution
Figure 1 Variable resolution caused by walls at different
distances from the scanner.
The above complicate the selection of optimum thresholds for
segmentation criteria. One possible solution to this problem is
to apply different thresholds optimised for local point cloud
resolutions. Alternatively, a single threshold can be used and
local segmentation criteria can be weighted based on local point
cloud resolution.
4. PROPOSED DETECTION ALGORITHM
The proposed solution is based on weighted proximity
segmentation. The method works on the assumption that the
mortar channel is reasonably deep and wide (at least two points).
The method is designed to work on a 3D point cloud. This is
necessary because the point cloud of a wall is typically
composed of more than one scan. The sequence of steps in the
algorithm is outlined in the following subsections.
4.1 Pre-processing
The rgb value of all points is converted to hsv. As mentioned in
section 3.5 this is done to untangle the effect of surface lighting
from the hue.
4.2 Triangulation
The point cloud is triangulated using a 3D Delaunay
triangulation, see Figure 2(b). The triangulation is stored in a
graph G(V, E). The points in the point cloud are given by the
set V and the edges of the triangles by the set E. The graph is
used as a data structure to search point neighbourhoods and
effecting the connected components (explained in section 4.5)
4.3 Point attributes
Radiometric and geometric attributes are associated with each
point. The radiometric attributes include the intensity, hsv
triplets, and functions of the intensity and hsv of neighboring
points. Geometric attributes include the statistics of the
Euclidean distance to the N nearest points, and the normal at a
point.
4.4 Edge attributes
From the point attributes edge attributes are determined. These
attributes represent the strength of the connection between
568