A NOVEL IMAGE ACQUISITION AND PROCESSING PROCEDURE FOR FAST TUNNEL DSM PRODUCTION

In mining operations the evaluation of the stability condition of the excavated front are critic to ensure a safe and correct planning of the subsequent activities. The procedure currently used to this aim has some shortcomings: safety for the geologist, completeness of data collection and objective documentation of the results. In the last decade it has been shown that the geostructural parameters necessary to the stability analysis can be derived from high resolution digital surface models (DSM) of rock faces. With the objective to overcome the limitation of the traditional survey and to minimize data capture times, so reducing delays on mining site operations, a photogrammetric system to generate high resolution DSM of tunnels has been realized. A fast, effective and complete data capture method has been developed and the orientation and restitution phases have been largely automated. The survey operations take no more than required to the traditional ones; no additional topographic measurements other than those available are required. To make the data processing fast and economic our Structure from Motion procedure has been slightly modified to adapt to the peculiar block geometry while, the DSM of the tunnel is created using automatic image correlation techniques. The geomechanical data are sampled on the DSM, by using the acquired images in a GUI and a segmentation procedure to select discontinuity planes. To allow an easier and faster identification of relevant features of the surface of the tunnel, using again an automatic procedure, an orthophoto of the tunnel is produced. A case study where a tunnel section of ca. 130 m has been surveyed is presented.


INTRODUCTION 1.1 Geostructural data acquisition in tunnels
The mining operations in an underground quarry run continuously all over the day and are organized in cycles each made by three phases: the perforation and the insertion of the explosive charges at the front, the detonation of the charges and the removal of the material.Since the drilling machine is guided by a laser system, points are marked along the walls of the tunnel and surveyed with a total station so that any subsequent operation can be referred to the tunnel reference system, monitoring the progress of the excavation and checking its conformity to plans.Technical prescriptions foresee that a geological report is produced which describes the general geostructural characteristics found during the progress of excavations.Periodically, a geological survey is therefore executed to verify the stability conditions at the front.The ISRM (International Society for Rock Mechanics) recommendations prescribe the sampling of all the discontinuities traces that intersect a certain reference line (scan line) or that fall within a certain window: for every discontinuity track detected the orientation is assessed using a geological compass and the spacing with respect to traces of the same family is measured with a tape and reported in a sketch.The gathered data are analysed generally using a statistical approach: by clustering all the spatial orientations of the discontinuities, their families are highlighted and the behaviour in case of collapse of the rock structure predicted.It is therefore essential that the number of discontinuities sampled is high enough to make the statistical analysis significant.Moreover, the geologist must evaluate the lithology of the rock mass, the presence of water and/or gas and carry out tests of resistance to compression.
Though routinely applied, this method has some drawbacks.Apart from the risk of falling stones in an area not yet consolidated, the measurement are restricted to the lower part of the tunnel, directly accessible by the geologist.The survey takes some time to carry out, which interrupts the excavation activities.Moreover, if the tunnel is intended for civil use (e.g. a road or a railway), the documentation gathered by the geologists cannot be verified later.In case of an accident due to rock fall during the construction or other damages occurring after the opening to traffic it's almost impossible to independently assess the actual initial geomorphological conditions and find out whether some design or construction error occurred: in other words it is not possible to verify the correspondence of the measured data with the previous state of the tunnel.Improving the data collection in terms of safety, amount and distribution of measurements without increasing or perhaps even reducing the time needed would improve the efficiency of the quarry management; a better documentation of the surveys would also be desirable.To improve the data acquisition technique, overcoming its current critical aspect, it was decided to develop an innovative system based on photogrammetry, where the geostructural characteristics are derived from measurement made on a high resolution DSM (Digital Surface Model).

Non-contact measurement techniques
So far the use of photogrammetric techniques in tunnels has been concentrated on the monitoring of the deformations and displacements in the period following the excavation, measuring periodically the three-dimensional displacements of targets applied to the vault and to the side walls (Nakai et al., 2003).The same operation are today mostly performed using a Terrestrial Laser Scanner (TLS) (Nuttens, T. et al., 2010;Fekete et al., 2009)   all images acquired from the same stop of the rotation device from the different stations along the tunnel axis.Once the recently excavated section of the tunnel is surveyed, the excavation front is also surveyed by a minimum of three images (figure 1.c).
Once a new section is cleared after the blast, reference marks are painted on the piers and surveyed to check the progress of the drilling machine according to plan.They can easily be recognized in the image sequence and therefore used as control points for the orientation of the block of strips.This allows the progressive georeferencing of blocks along the tunnel, consistent with the tunnel reference system, without additional survey work that may interfere with the excavation.

Image block design and simulation
In a preliminary stage a simulation has been carried out to test the feasibility of the method.A tunnel section 10 m long, 8 m wide and 7.2 m high with rough surfaces has been generated in a virtual environment and draped with a photorealistic texture.This is roughly the amount of advance after each blast in the mining tunnel under study.Finally, artificial illumination reproducing the one available on site has been introduced.The synthetic block is made of three transversal strips with longitudinal baselenght of 2.1 m; the survey of piers and vault is completed by 6 images spaced by 36°.Overall, therefore, the block is made of 6 longitudinal strips of 2 models each, plus 3 images of the front.Synthetic images were generated and processed according to the pipeline described in the next section; an artificial DSM has been generated by image correlation.The results, though clearly obtained under ideal conditions, allowed to draw useful hints for the realization of the panoramic head and the strong influence of rock texture and illumination on the outcome.

Image block orientation
Many approaches have been developed in the past years in the photogrammetric and CV communities for the automated orientation of large and complex image sequences.The former discipline focuses primarily on precision and reliability of the results (Pierrot-Deseilligny et al., 2011).The latter is more concentrated in the orientation of large blocks (thousands of images) taken with un-calibrated cameras without any metric purpose except 3D visualization and image browsing (Snavely et al., 2008).
The main problems preventing the presence of reliable and precise commercial approaches are the presence of convergent images, unpredictable baselines and scale variations, lighting changes, repetitive patterns, homogeneous textured areas, etc.In this case, however, the geometric configuration of the block is quite regular: the base-length between consecutive images are always approximately the same and can be contained to grant a good overlap along the sequence, helping the identification of good tie points; all the images have a normal pose with respect to the object, basically analogous of that in an aerial block; the object usually presents, at this image scale, very good and contrasted elements that makes the tie point reconnaissance easy and very fast.
Consequently the automatic orientation approach has been designed based on simple SfM algorithms already developed by our research group for other, more demanding, applications: the orientation strategy follows a simple Feature Based Matching (FBM) procedure at the end of which the operator has only to identify the Ground Control Points (GCP) needed to georeference the image block.
Most FBM techniques used for image orientation are based on the use of detector/descriptor operators.The detector is capable of finding interest points in the images, while the descriptor associates a vector of information to each single detected point.As shown in (Roncella et al., 2011) the use of a scale-invariant feature extractor/descriptor, in this particular framework, is essential to obtain good results: as far as straight tunnel section are considered the use of any interest operator/descriptor is unimportant; due to the high overlap between consecutive images, and thanks to the high regularity in the block geometry and very good image texture, any strategy reach very good results.Nonetheless, sometimes, on the side walls of the tunnel apertures and intersection with other galleries are present: in that case the perspective changes between consecutive images are much higher and a non-scale invariant operator can lead to unpredictable results.
For this reason a SURF operator/descriptor has been implemented in the FBM routine: the SURF operator is partly derived from the SIFT detector and uses an integer approximation to the determinant of the Hessian, which can be computed very fast with integral images: while the SIFT operator is computationally very heavy, the SURF operator has a computational load comparable to that of the Harris operator (commonly used in many SfM strategies, even if it's only rotation invariant).Moreover, its scale invariance properties should grant higher repeatability compared to Harris.
To select the corresponding features by comparing directly the values of the descriptors, an exhaustive search along the lists can be performed; the computational load can be very high, though, due to the large number of features usually extracted by the operator.It is more efficient to use approximate methods like the kd-tree (Beis et al., 1997), that are faster but still deliver extremely reliable results.
As far as the epipolar geometry computation is concerned, rather than estimate the fundamental matrix, the 5-point calibrated relative orientation algorithm proposed by Nister (2004) has been implemented.Compared to the fundamental matrix, it offers a more straightforward approach that exploits the fact that, in practice, in all photogrammetric surveys the interior orientation elements (as well as image distortion parameters) are known.The interior orientation is implicitly forced into the estimation using normalized image coordinates and obtaining directly the essential matrix.The mathematical model is more complex (it uses a 10 th degree polynomial instead of the 3 rd degree one for the fundamental matrix estimation); nonetheless, since it uses a more correct deterministic model, the estimation with the RANSAC algorithm achieves a higher inlier percentage: being the iterations dependent on the inlier percentage, the method is computationally more efficient than the old one.
We decided also to enforce the interior orientation during the three view geometry estimation.In this case it's much harder to implement the interior orientation parameters inside the trifocal tensor estimation: instead of using the trifocal tensor (Hartley et al., 2000) a RANSAC bundle block adjustment using just three images is used: six points are extracted randomly from the correspondences filtered out in the previous stages and the relative orientations parameters between a reference image and each of the two others images are computed; using the 6 points the three images are referred to the same object space; then, fixing the orientation parameters, all the other point correspondences are triangulated and image projection residuals are computed: all the points with a mean projection residual lower than a specified threshold are considered inlier.Maybe it can be less time consuming a strategy where the orientation parameters co directly used: object space stage is usually dramatically lo orientation sol At the same reconstruction afterwards.
At the end of t acquired along link between a the particular photo-station: photo-station axes that pass itself; the tra frames acquir homography t images: Where H is coordinates) re K is the calibr interior orient matrix betwee Since the rotat the acquisition are known, th between them tries to link th other image: t between the tracked for at points extracte approach, the on one sequen (LSM, Gruen, At the end of with all the operator) is pe

DSM gen
The

Rockscan
The As mentioned in Section 3.1, exploiting the approximate homography between strips to connect them in a single block, the points extracted in the overlap area of two longitudinal strips are transferred to the adjacent one to tie the strips together.
After the registration of all the images with a bundle adjustment and the successive DSM generation, a detailed orthophoto of the internal surface of the tunnel was created; the orthophoto relative to a section of about 40 m is shown in figure 4.

CONCLUSIONS
The technical feasibility of performing geostructural surveys where discontinuity planes and traces are measured on high resolution DSM and oriented images has been shown for some time.The challenge is today to find technical solutions that allows the method to be applied advantageously in as many cases as possible.The results illustrated in this paper show that an effective operational pipeline can be set that allows minimal disruption of the mining operations, though guaranteeing correct results from the geostructural survey.In addition, consistent georeferencing and documentation with oriented images foresees the progressively builds up of a valuable database.Further progress in data acquisition speed and safety of operation might be achieved by installing the camera and the device on the drilling machine, though this require a specific remote controlling by actuators of the rotating device and of the camera.
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XXXIX-B5, 2012 XXII ISPRS Congress, 25 August -01 September 2012, Melbourne, Australia FigurOrthophoto p he end of the D plified to be a ogists.To this a reprojected to g tunnel.Since plex, a simplifi mapping of it posed of two p ch support a se t under the u ogonally on th e the DSM p ments are pro ular vault.

Figure
Figure 3. S oftware to prod ery was proprie all the DSM p metry: the dime International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XXXIX-B5, 2012 XXII ISPRS Congress, 25 August -01 September 2012, Melbourne, Australia