ISPRS / CIPA Workshop «UNDERWATER 3D RECORDING & MODELING» 16 17 April 2015 Piano di Sorrento (Napoli), Italy Building a 3D reference model for canal tunnel surveying using SONAR and LASER scanning E. Moisan, P. Charbonnier, P. Foucher : Cerema DTer Est P. Grussenmeyer, S. Guillemin, M. Koehl : INSA Strasbourg Direction territoriale Est 1
The project 33 canal-tunnels in France (42km) Built during the 19th and the 20th centuries Conserving this heritage Protecting goods and persons Periodical inspections are necessary Devising an automatic inspection system 2
Objectives Since 2009 Developement of a dynamic acquisition system for the abovewater part using photogrammetry Addition of an underwater recording system Accuracy assessment a reference model Static acquisitions along the tunnel 2013: above water reference model (accuracy 2 cm) Build full 3D Model 3
Constraints No GPS Signal Water turbidity Confined environment AOther location methods ASonar technology AAdapted methods 4
Selected methods Above-water part 3D Terrestrial LASER Scanner TLS Faro Focus 3D Underwater part 3D Mechanical Scanning SONAR MSS Blueview BV5000 Noisy data Robust methods Lack of overlap link both environments 5
Selected methods Above-water part 3D Terrestrial LASER Scanner TLS Faro Focus 3D Underwater part 3D Mechanical Scanning SONAR MSS Blueview BV5000 Noisy data Robust methods Lack of overlap link both environments a Using ladders 6
Experimental site Niderviller canal tunnel Built 1839-1845 475m long acquisition setup (top view) 2 MSS Stations 2 TLS Stations 2 ladders spheres 7
Experimental site Niderviller canal tunnel Built 1839-1845 475m long acquisition setup (top view) 2 MSS Stations 2 TLS Stations 2 ladders spheres 8
Experimental site Niderviller canal tunnel Built 1839-1845 475m long acquisition setup (top view) 2 MSS Stations 2 TLS Stations 2 ladders spheres 9
Devices Data recording 3D surface reconstruction Registration & geo-referencing Results T L S Horizontal resolution 0.036 (3 mm/10m) Vertical resolution 0.036 (3mm/10m) Beam width 2.25mm + 2x0.011 Ranging error 2 mm (10-25 m) Maximum range 330m Field-of-view 300 /360 Horizontal resolution Vertical resolution ~0.069 (16mm/10m) 0.18 (30mm/10m) M S S Beam width 1 /1 Ranging error 15 mm Maximum range 30m Field-of-view 45 /360 (320 /360 ) 10
Sensors Data recording 3D surface reconstruction Registration & geo-referencing Results T L S Horizontal resolution 0.036 (3 mm/10m) Vertical resolution 0.036 (3mm/10m) Beam width 2.25mm + 2x0.011 Ranging error 2 mm (10-25 m) Maximum range 330m Field-of-view 300 /360 Horizontal resolution Vertical resolution ~0.069 (16mm/10m) 0.18 (30mm/10m) M S S Beam width 1 /1 Ranging error 15 mm Maximum range 30m Field-of-view 45 /360 (320 /360 ) 11
SONAR resolution Multi-beam echo-sounder swaths under two different incidence angles Acquisition context and MSS capacities impact imagery resolution Estimation of SONAR horizontal resolution 12
Denoising Why? Using meshing to denoise SONAR data Helps interpretation highlights errors Guides denoising Use all cloud points: many meshing artifacts 13
Meshing methods Two methods Use cloud points as vertices At a large scale : Denoising J Loss details L Compute nearest surface by interpolation of new points Denoising J Oversmooth L 14
Meshing methods Coarse-to-fine method Increase mesh definition Trade-off between details and noise interpretation Drawbacks : several manual operations + lots of interpretation 15
Registration methods Direct Device on known coordinates point Surveying the position Based on targets spheres TLS ladders MSS Indirect Based on point clouds ICP MSS co-registration using geometrical entities TLS/MSS registration 16
Methods of registration Direct Device on known coordinates point Surveying the position Based on targets spheres TLS ladders MSS Indirect Based on point clouds ICP MSS co-registration using geometrical entities TLS/MSS registration 17
Underwater geo-referencing Get geo-referencing by registration of MSS data on TLS data Main issue : lack of overlap between TLS and MSS acquisitions Attitude correction Vertical translation Horizontal translation Procrustes Ladders Waterline 18
Attitude correction Geometrical entities can approximate parts of above and underwater model. canal banks planes salient elements lines Uses orthogonal Procrustes 19
Vertical translation Ladders are used to compute the difference in altitude 3-step robust method to estimate ladders in the noisy underwater point cloud 1 Manual segmentation of ladder points in point cloud Estimation of the ladder plane Projection of points on the ladder plane 2D Data 20
Vertical translation 2 Automatic segmentation of ladder components A Histogram based detection 0.5 m 21
Vertical translation 3 Robust adjustement of straight lines Takes into account parallelism and orthogonality Orthogonal fit (residuals) Simultaneaous robust estimation based on M-estimators 22
Vertical translation 3 Robust adjustement of straight lines Takes into account parallelism and orthogonality Orthogonal fit (residuals) Simultaneaous robust estimation based on M-estimators 23
Horizontal translation 1. Waterline 2D silhouette extraction on the TLS and MSS model 2. 2D ICP algorithm application to estimate translation vector 24
Horizontal translation 1. Waterline 2D silhouette extraction on the TLS and MSS model 2. 2D ICP algorithm application to estimate translation vector 25
Obtained model M es h P oi n ts 26
Discussion Method to record and process TLS and MSS data First experience of underwater acquisition in canal tunnel Acquisition setup SONAR capacities + confined environment difficulties to co-register MSS data decrease distance between stations use more targets (e.g. ladders) 27
Discussion Denoising by meshing Noisy nature of SONAR data many manual operations lots of interpretation Other denoising techniques can be explored Full 3D Reference model? limits of the technology and knowledge accuracy assessment Experiments in controled conditions (e.g. dry dock) 28
Direction territoriale Est Thank you for your attention Emmanuel MOISAN Phd Student +33 (0)3 88 77 79 17 emmanuel.moisan@cerema.fr ISPRS / CIPA Workshop «UNDERWATER 3D RECORDING & MODELING» 16 17 April 2015 Piano di Sorrento (Napoli), Italy
ISPRS / CIPA Workshop «UNDERWATER 3D RECORDING & MODELING» 16 17 April 2015 Piano di Sorrento (Napoli), Italy Building a 3D reference model for canal tunnel surveying using SONAR and LASER scanning E. Moisan, P. Charbonnier, P. Foucher : Cerema DTer Est P. Grussenmeyer, S. Guillemin, M. Koehl : INSA Strasbourg Direction territoriale Est 30
Solution to orthogonal Procrustes problem The orthogonality property must by taken into account in the computation of the rotation matrix (orthogonal Procrustes problem) min ǁ A BQ ǁ F ² subject to Q T Q=I Solution : compute the Singular Value Decomposition of B T A U T (B T A)=Σ Q=UV T 31
M-estimator Replace the usual sum of squared residuals by a function of the form : ρ is a non quadratic potential Minimize J(θ) is equivalent to minimize : Weighted least square The minimum is obtained for : Weighting function of residuals Algorithm: Iterated Reweighted Least Squares (IRLS) 32