117-124 Poropat

The South African Institute of Mining and Metallurgy
International Symposium on Stability of Rock Slopes in Open Pit Mining and Civil Engineering
Marc Elmouttie, George Poropat
AUTOMATED STRUCTURE MAPPING OF ROCK FACES
G.V. Poropat & M.K. Elmouttie

CSIRO Exploration & Mining, Brisbane, Australia
ABSTRACT
Mapping of exposed rock walls provides data on geological features such as joints,
bedding planes, faults etc critical to the economics and the safety of mine operations.
Traditional, manual methods of mapping can be time consuming, error prone and limited
in their coverage. New methods have been developed to map exposed faces remotely and
safely however the benefits of these methods are only fully realised when exposed
structures can be quickly and easily mapped using automated methods.
This paper describes the application of an automated method of analysis of 3D images that
provides users with the capacity to map exposed structures automatically. The method is
based on the use of vector classification methods used in some neural network systems and
provides rapid and reliable identification of exposed planes. CSIRO Exploration and
Mining is applying these methods to 3D images for a range of applications from pit
mapping to machine automation.
1 INTRODUCTION
Knowledge of geological features such as faults, bedding planes and joints reveals
information pertaining to the economics and the safety of the mine operation. There are
several methods to map the structure of a rock wall. Traditionally, mapping has been
accomplished with manual measurement techniques such as those described in Priest &
Hudson (1981). These procedures are time consuming, possibly dangerous and the
opportunity for errors to occur during measurement and data entry phases is considerable.
More recently, laser range finding (Feng & Roshoff, 2004) and 3D photogrammetry
(Poropat, 2001) have been employed. Both techniques offer superior accuracy and
efficiency over manual measurements. Laser ranging methods acquire data as point clouds
but require expensive equipment. 3D photogrammetry can be accomplished with a great
deal of flexibility using low-cost digital camera technology.
No matter what technique is used to undertake the measurements, the large amounts of
data obtained in mapping the surface topography of a rock mass require efficient and
reliable processing methods to extract structure information such as the joint set
parameters. Ideally, these methods should be automated in order to deal with large or
numerous data sets and reduce the influence of measurement bias. This requirement drives
the need for a suitably reliable pattern recognition algorithm.
Neural networks of various architectures have been used for a wide variety of pattern
recognition, exploratory data analysis and visualization applications (Seiffert & Jain,
2002). The ability of neural networks to quantify and classify vectors of any dimension
reduces a multi-dimensional problem into a more manageable low-dimensional (e.g. 2D)
one.
Page 117
This paper describes the use of such techniques to map the structure of rock-walls. The
methodologies used and some examples of results are presented below.
2 IDENTIFICATION OF PLANES
2.1 Pre-processing The Input Data
The method we used to acquire stereo images and generate the 3-D data is detailed in
Soole & Poropat (2000). In short, using stereo photogrammetry, a number of rock-walls
were imaged using the Sirovision system. The system images the subject walls from two
defined viewing positions producing 3D spatial coordinates. A triangular mesh
representation of the surface is then created to support calculation of surface properties.
The effectiveness of the neural network can be increased by incorporating both surface
orientation data and measures of ‘flatness’. These surface properties can be determined
computationally using the 3D mesh (Flynn & Jain, 1989). To ensure their accuracy,
suitable edge-sensitive smoothing must be applied to the mesh prior to computation. In
particular, the flatness data assists in detection of non-planar structures which nonetheless
possess regular morphologies (e.g. curved edges, bowls etc).
2.2 The Structure Analysis Algorithms

The power of neural networks comes from their ability to accurately scale the topology
and distribution of the multi-variate input data space into a conveniently low-dimension
space which is more amenable to well established data processing techniques. A clustering
algorithm is used to sort the output of the neural network. As the goal of the structural
analysis is essentially to detect regular surface types such as planes or edges (hereafter
referred to as features) in the data, the number of clusters in the data is directly related to
the number of types of such structures exposed on the rock surface.
Automatic determination of the cluster number is possible, although it is often desirable to
limit the search to a particular number of prominent joint sets. Visualisation of each of the
detected features will reveal spatially separated joints, bedding planes etc. These features
have been clustered together based on their similarity as measured using the surface
normal and flatness parameters. However, each feature can be labelled individually using
image processing algorithms (Haralick & Shapiro, 1992) and thus surface properties (i.e.
area, orientation, average flatness, perimeter etc) can be ascribed to each. Thus the
structure analysis reveals information about joint sets and individual joints, edge sets and
individual edges, loose material and other features.
3 RESULTS AND DISCUSSION

3.1 Source Data
Analysis was performed on Sirovision data acquired at a rock wall. A 6 m x 17 m section
of the wall was imaged and is shown in Figure 1. Several distinct features such as joint sets
and bedding planes are present as well as some loose material at the base of the wall.
Page 118
Figure 1 The rock wall

The structure detection algorithm successfully identified the major joint faces and several
bedding planes present in the rock wall and these are shown in Figure 2. Both large and
small planes have been detected, including planes which are nearly parallel to the line of
sight. Critically, non-planar regions of similar morphology are also detected by this
algorithm, such as the loose material at the base of the wall.
Page 119
Figure 2 The major structures detected on the 3D surface

Figure 3 presents a particular region of interest containing two dominant planes separated
by a curved, rounded edge. The results illustrate the ability of the algorithm to distinguish
surface regions with complex (non-planar) features from more regular, planar structures.
The figure clearly shows successful detection of the boundary between the offset planes.
Page 120
Figure 3 The offset plane region of interest (top) and the algorithm results.
Although full automation is possible, there is some manual control over processing that
still can be provided by the user. Filtering based on joint areas can eliminate unwanted
detection of small surface irregularities. User control of process tolerance when using
Page 121
orientation or curvature data can improve both processing efficiency and accuracy.
However, meaningful results are possible with no user input (as shown in these figures).
3.2 Structural Mapping In Open Pit Operations

The structure analysis was applied to a 3D model of an open pit as shown in Figure 4. This
model represents a much larger surface than the previous example. Further, the resolution
of this model is such that small-scale features have been suppressed. Nonetheless, the
results in Figure 5 show the algorithm’s ability to clearly separate out the bench faces and
berms. Further, discrimination of the dominant planes in the face data is also seen.
Figure 4 A 3D image of a section of an open pit
Page 122
Figure 5 Feature detection using the automated algorithm.

If the number of clusters is (manually) set to 5, only dominant features are detected by the
algorithm (Figure 6). Limiting the search to large-scale features produces similar results.
Figure 6 Manually limiting the algorithm’s search to 5 clusters
Page 123
4 CONCLUDING REMARKS
Automated structure mapping of 3-D spatial data has been demonstrated. Using a neural
network as the pattern recognition tool, with pre- and post-processing of the data involving
surface smoothing algorithms and image processing techniques, joint sets and individual
joints have been identified and surface properties measured automatically. Future work
will involve the implementation of more sophisticated edge detection algorithms to assist
the pattern recognition phase of the processing.
REFERENCES
PRIEST S.D. & HUDSON J.A. 1981 Estimating Discontinuity Spacing and Trace Length
Using Scanline Surveys Int J. Rock Mech. Min. Sci. & Geomechanics 18: 183 – 197
FENG Q.H. ROSHOFF K. 2004 Int J. Rock Mech. Min. Sci. 41: 379 – 384
POROPAT G. New methods for mapping the structure of rack masses EXPLO 2001 Conf
Proc.
SEIFFERT U. & JAIN L.C. 2002 Review of "Self-organizing neural networks: Recent
advances and applications" Physica-Verlag, Heidelberg
SOOLE P. & POROPAT G. 2000 Mine site mapping using Terrestrial Photogrammetry.
Bowen Basin Symposium, Rockhampton, Australia
FLYNN, P.J. & JAIN, A.K. 1989 On Reliable Curvature Estimation. In Computer Vision
and Pattern Recognition. Proceedings CVPR '89. IEEE Computer Society Conference on,
pp 110-116
HARALICK R. M. & SHAPIRO L.G. 1992 Computer and Robot Vision, Volume I,
Addison-Wesley, pp. 28-48.
Page 124

117-124 Poropat

Uploaded by

Copyright:

117-124 Poropat

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

117-124 Poropat

Uploaded by

Copyright:

The South African Institute of Mining and Metallurgy

AUTOMATED STRUCTURE MAPPING OF ROCK FACES

G.V. Poropat & M.K. Elmouttie

2.2 The Structure Analysis Algorithms

3 RESULTS AND DISCUSSION

Figure 1 The rock wall

Figure 2 The major structures detected on the 3D surface

3.2 Structural Mapping In Open Pit Operations

Figure 4 A 3D image of a section of an open pit

Figure 5 Feature detection using the automated algorithm.

Figure 6 Manually limiting the algorithm’s search to 5 clusters

You might also like