Accuracy assessment of a canal-tunnel 3D model...

1
Accuracy assessment of a canal-tunnel 3D model by comparing photogrammetry and laserscanning recording techniques P. Charbonnier * , P. Chavant * , ** , P. Foucher * , V. Muzet * , D. Prybyla * , T. Perrin * , P.Grussenmeyer ** , S. Guillemin ** * : CETE de l’Est, LRPC de Strasbourg, 11 rue Jean Mentelin, 67035 Strasbourg, France. {pierre.charbonnier, philippe.foucher, valerie.muzet, davy.prybyla, thibaut.perrin}@developpement-durable.gouv.fr ** : ICube laboratory UMR 7537, Photogrammetry and Geomatic Group, INSA de Strasbourg, 24 bd de la Victoire, 67084 Strasbourg, France. {paul.chavant, pierre.grussenmeyer, samuel.guillemin}@insa-strasbourg.fr 1. CONTEXT •Design of a visual inspection system •Test in a 475 m long canal tunnel • Images Photogrammetric model • Terrestrial Laser Scanning • Control points for georeferencing Figure 1: Terrestrial Scanner Laser C10 and resulting 3-D model (TLS) Figure 2: Photogrammetric model (PM), camera configuration and obtained model (with camera poses) 2. COMPARISON METHODS Terrestrial Laser Scanning (TLS) Static acquisition: • 15 stations (1h per station) • Cloud absolute referencing • Scan strip registration • 1.5 billion points Accuracy assessment of PM model taking TLS as a reference Photogrammetric model Dynamic acquisition: • More than 500*2 images • 3-D reconstruction : 334 900 points • 9 submodels Accuracy assessment distance between a set of points and a reference set. Huge and heterogenous point cloud Data resampling and specific computation structure Distance computation 3 approaches : (1) Point-to-point method: point to point distances computation (2) Cloud-to-triangulated mesh method: euclidean distances between point cloud and mesh (Delaunay triangulation) X Y Z 952635.231 124890.738 270.068 952453.653 124748.478 270.043 952409.288 124713.742 270.053 X Y Z 952635.239 124890.735 270.056 952453.646 124748.477 270.024 952409.297 124713.747 270.036 0.014 m 230 m 287 m (3) Cloud-to-cloud method: Octree structure, surface modeling, two step filter (Hausdorff distance), euclidean distances computation 3. RESULTS Datasets PMS TLS Length 1 6 scattered in tunnel 2 12507 68429 18 m 3 63295 76578 107 m 4 226679 201838 whole tunnel Table 1: Number of points in both models for each dataset Datasets Computation time Accuracy 2 < 1 mn 5.7 cm 3 3.7 mn 5.8 cm 4 24 mn 4.1 cm Table 2: “Cloud-to-triangular mesh” method: Accuracy comparison between photogrammetric and TLS model Datasets Computation time Accuracy 2 3.37 s 4.1 cm 3 22.32 s 5.6 cm 4 8.6 mn 4.2 cm Table 3: “Cloud-to-cloud” method: Accuracy comparison between photogrammetric and TLS model. Seven octree levels are used Dataset 1 Leica geosystems spherical targets From TLS model and tacheometry Accuracy 1.7 cm Result TLS is a reference 4. CONCLUSION • Several methods for point cloud comparison in 3-D Speed constraints: Octree subdivision appropriate resampling TLS model as reference • Photogrammetric model accuracy : around 5 cm

Transcript of Accuracy assessment of a canal-tunnel 3D model...

Page 1: Accuracy assessment of a canal-tunnel 3D model …eprints2.insa-strasbourg.fr/1485/3/Poster_CHAVANT_Paul.pdfmodel by comparing photogrammetry and laserscanning recording techniques

Accuracy assessment of a canal-tunnel 3D model by comparing photogrammetry and

laserscanning recording techniques P. Charbonnier*, P. Chavant*,** , P. Foucher*, V. Muzet*, D. Prybyla*, T. Perrin*, P.Grussenmeyer**, S. Guillemin**

* : CETE de l’Est, LRPC de Strasbourg, 11 rue Jean Mentelin, 67035 Strasbourg, France. {pierre.charbonnier, philippe.foucher, valerie.muzet, davy.prybyla, thibaut.perrin}@developpement-durable.gouv.fr

** : ICube laboratory UMR 7537, Photogrammetry and Geomatic Group, INSA de Strasbourg, 24 bd de la Victoire, 67084 Strasbourg, France. {paul.chavant, pierre.grussenmeyer, samuel.guillemin}@insa-strasbourg.fr

1. CONTEXT • Design of a visual inspection system • Test in a 475 m long canal tunnel •  Images à Photogrammetric model •  Terrestrial Laser Scanning •  Control points for georeferencing

Figure 1: Terrestrial Scanner Laser C10 and resulting 3-D model (TLS) Figure 2: Photogrammetric model (PM),

camera configuration and obtained model (with camera poses)

2. COMPARISON METHODS

Terrestrial Laser Scanning (TLS)

Static acquisition: •  15 stations (1h per station) •  Cloud absolute referencing •  Scan strip registration •  1.5 billion points

Accuracy assessment of PM model taking TLS as a reference

Photogrammetric model

Dynamic acquisition: •  More than 500*2 images •  3-D reconstruction : 334 900 points •  9 submodels

•  Accuracy assessment Ø  distance between a set of points and a reference set.

•  Huge and heterogenous point cloud Ø  Data resampling and specific computation structure

•  Distance computation à 3 approaches : (1) Point-to-point method: point to point distances computation

(2) Cloud-to-triangulated mesh method: euclidean distances between point cloud and mesh (Delaunay triangulation)

X Y Z

952635.231 124890.738 270.068 952453.653 124748.478 270.043 952409.288 124713.742 270.053

X Y Z

952635.239 124890.735 270.056 952453.646 124748.477 270.024

952409.297 124713.747 270.036

0.014 m

230 m 287 m

(3) Cloud-to-cloud method: Octree structure, surface modeling, two step filter (Hausdorff distance), euclidean distances computation

3. RESULTS Datasets PMS TLS Length

1 6 scattered in tunnel 2 12507 68429 18 m 3 63295 76578 107 m 4 226679 201838 whole tunnel

Table 1: Number of points in both models for each dataset

Datasets Computation time

Accuracy

2 < 1 mn 5.7 cm 3 3.7 mn 5.8 cm 4 24 mn 4.1 cm

Table 2: “Cloud-to-triangular mesh” method: Accuracy comparison between photogrammetric and TLS model

Datasets Computation time

Accuracy

2 3.37 s 4.1 cm 3 22.32 s 5.6 cm 4 8.6 mn 4.2 cm

Table 3: “Cloud-to-cloud” method: Accuracy comparison between photogrammetric and TLS model.

Seven octree levels are used

Dataset 1

•  Leica geosystems spherical targets

From

•  TLS model and tacheometry

Accuracy •  1.7 cm

Result •  TLS is a reference

4. CONCLUSION • Several methods for point cloud comparison in 3-D •  Speed constraints:

•  Octree subdivision •  appropriate resampling

•  TLS model as reference • Photogrammetric model accuracy : around 5 cm