The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Voi. XXXVII. Part B4. Beijing 2008
building edges. In applications seeking to determine changes
between two or more epochs of data over an urban or peri-urban
area, where the demolition or construction of
buildings/structures is of interest, lidar data, in contrast with
aerial photography, is potentially able to provide a point cloud
dataset requiring significantly less manual editing than that
produced by conventional photogrammetric processes.
However, while the efficiency of data capture might be
improved, it is also necessary to improve the methodology
relating to change detection in order to cope with this potential
for an increase of data. Techniques operating on the point
clouds themselves also have the advantage of being able to be
applied to point clouds from any source, making such an
approach more versatile. Efficient change detection
methodologies, coupled with increasingly frequent data capture
could lead to a merging of continuous and cyclic revision
processes. Even without an increase in the quantity or scale of
data it would be of interest to provide robust routines that can
identify areas of interest automatically.
1.2 Point cloud based change detection
As high resolution three-dimensional samples of the real world,
point clouds have potential for change detection at a variety of
scales. Change detection on the point clouds themselves would
help to increase the speed of the change detection process and
improve the management of the datasets (both key parameters
in determining the cost efficiency of the change detection
process). There are, however, a number of obstacles to the
demonstration of a robust change detection process based solely
on a three-dimensional point cloud:
• ‘best guess’ point correspondence must be determined, even
though no direct point to point correspondence can be
assumed;
• known topology cannot be assumed as point clouds may
originate from multiple sources;
• occlusions during data collection may need to be considered;
• differences between two point clouds do not imply an
information change (e.g. the addition of a new building);
• large datasets require memory-efficient data handling
routines.
The development of change detection procedures that operate
on point cloud datasets would allow national mapping agencies
to take advantage of improved data flow lines. As a result they
would be able to provide change detection at a higher temporal
and spatial scale, or simply be able to improve the economics of
map revision.
This paper will outline work to establish a flow line to identify
change between two point cloud datasets, based on the
comparison of octrees, i.e. three-dimensional data structures
representing the point cloud. By basing any comparison on
these octrees, it is hoped to limit the amount of pre-processing
required (for example the building of a digital surface model).
The paper will start with a brief introduction to octrees. It will
then present results of using the methodology on a variety of
datasets. This will include datasets derived from digital aerial
photography. Finally it will provide a discussion on the
performance of the system and outline some conclusions.
2. METHODOLOGY
2.1 Octrees
An octree (Figure 1) is a recursive and regular subdivision of
three-dimensional space originally used to optimise computer
graphics rendering (Botsch at al., 2002). The bounding box for a
three-dimensional geometry is divided into eight equal cubes,
each of which are in turn divided into eight more cubes until a
specified level of subdivision is achieved, or the cubes have less
than a predefined number of data points within their bounds. At
this point leaf nodes are formed. An octree representation of a
point cloud provides an efficient represent of the spatial
distribution of the 3D points. The tree structure assumes no
previous knowledge of topology but allows efficient searching
of the point cloud to determine possible correspondence.
Figure 1 An octree generated from data used in the study
The comparison of point clouds against 3D CAD models is
currently undertaken using point-to-mesh and mesh-to-mesh
processes, often using algorithms based on the iterative closest
point solution outlined by Besl and McKay (1992). This
requires an existing reference model to be generated before
change detection can take place. Girardeau-Montanut et al.
(2005) outline a point-to-point comparison framework using
octrees to optimise the performance of the change detection
process, removing the need for a reference model; however this
process is based on point clouds collected from similar positions
and orientations and does not provide a general approach to the
problem. Nevertheless, octrees clearly have the potential for
facilitating the change detection directly between point clouds.
2.2 Node to node comparison
In this study the following methodology has been applied. Two
octrees are generated based on the two point clouds which are to
be compared. Each of these octrees share a common central
node (described in the same geographic coordinate system) and
the same controlling parameters, namely the maximum number
of points allowed in each node and the maximum number of
subdivisions allowed in the tree. As this study only deals with
point clouds which are defined in a common coordinate system
(for example the mapping system in use) one octree can be
compared with the other. If the point cloud datasets were
identical, the structure of each octree will be the same. Thus
differences in the structure of each octree can be termed as
change in the intervening time between epochs. Clearly, the
cause of this change must be taken into account as such change