development of a new algorithm for automated point ... · ² segregaon of small two-dimensional...

COSTissupportedbytheEUFrameworkProgrammeHorizon2020 TheFinalConferenceoftheCOSTAcDonTU1208,Warsaw,2017

Development of a new algorithm for automated point extraction from hyperbolic reflections in GPR data and

comparison with SPOT-GPR results

AleksandarRis-ć,MiroGovedarica,MilanVrtunski,ŽeljkoBugarinović(FacultyofTechnicalSciences,NoviSad,Serbia);LaraPajewski(SapienzaUniversity,Rome,Italy);XavierDerobert

(IFSTTAR,Nantes,France);SimoneMeschino(AirbusD&S,Germany)

AnexampleofcooperaDonbetweenscienDstsfromFrance,Germany,ItalyandSerbiawhostartedworkingtogetherthankstotheCOSTAcDonTU1208

TheFinalConferenceoftheCOSTAcDonTU1208,Warsaw,2017

§  SoluDons for fast andaccurate (automated)dataextracDon fromGPRdataareahighlyinteresDngtopicforGPRscienDficcommunity

§  Approaches:Signalprocessingorimageprocessing§  Imageprocessing–Dmedemanding,noisesensiDvi`y

Ø  Full,denseradargramsortresholded,sparseradargrams²  SimplificaDon – extracDon of data from hyperbolic reflecDons

(e.g.binarizaDon)²  SegregaDon of small two-dimensional secDons from a dense

radargram(segmentsofinterest-SOI)anddataextracDonfromthehyperbolicreflecDons

Ø  Unsupervised (e.g. Hough transform) or supervised procedures(e.g.,ArDficialNeuralNetworks–ANN)

Reasons & Related work

2


General steps of the automated algorithm

5.Dataextrac-on

4.Detectionandremovalofinterferedhyperbolas

2.Findingapexes 3.Findingpointsontheprongs

1.Detec-onofsegmentsofinterest(SOI)


§  STEP1:Conversiontorasterformatusingrad2bmp.§  STEP2:LocaDngofSOIusingMatLABGUItrainCascadeObjectDetector

(COD)Ø  Trainingset:posiDveandnegaDvetrainingsamplesØ  Usingrealand/orsyntheDcdata

1. Detection of segments of interest (SOI)

4


2. Finding apexes (slide 1/2)

§  STEP1:FindingtheiniDalpixelforsearchinSOI§  STEP2:Formingasub-matrixwithdimensions3x2aroundofiniDalpixel

withcurrentpixelinthesecondrowandthefirstcolumn§  STEP3:Findinglocalmaximumsinsub-matrix3x2(iteraDvely)

5


2. Finding apexes (slide 2/2)

§  Pixelswithminimumrowindexindicatetherowwhereapexislocated§  Points are divided into two groups: the ones that belong to the lef

prongandtheonesthatbelongtotherightprong§  Sr1 and Sr2 are calculated asmean values of pixels in the bordering

rowsofrightandlefprongs(yellowzone)§  ThecolumnoftheapexiscalculatedasthemeanvalueofSr1andSr2

6


3. Detection of points on the prongs

§  ThestartofthesearchisØ  Theupperrightcorner(forthelefprong)Ø  Theupperlefcorner(fortherightprong)

§  Finding local maximum in a sub-matrix 2x2, iteraDvely, to thestoppingcriterion(pixel intensityraDoofapexpointandcurrentsearchwindowpointsonprong)

7


4. Detection and removal of interfered hyperbolas

§  STEP1:Findingcrossingpoints(maximumdifferencebetweenrowandcolumn of the pixel at the posiDon of lef and right prong crossing isone)

§  STEP2:Checkingwhetherthecolumnsofcrossingpointscontainapexwhichislocatedunderthecrossingpoints

§  STEP3:Eliminateinterferedhyperbolas

8


§  WithminormodificaDons,thisalgorithmcanalsobeappliedtothecaseofdistrictheaDngsystems

Algorithm applications on complex structures

9


§  A.RisDć,Ž.Bugarinović,M.Vrtunski,M.Govedarica,“Pointcoordinateextrac-onfromlocalizedhyperbolicreflec-oninGPRdata”,JournalofAppliedGeophysics,Volume144,September2017,Pages1-17

§  A.RisDć,Ž.Bugarinović,M.Govedarica,L.Pajewski,X.Derobert,“Verifica-onofAlgorithmforPointExtrac-onfromHyperbolicReflec-onsinGPRData”,IWAGPR2017,9thInternaDonalWorkshop,Edinburgh,UK,June28-302017,DOI:10.1109/IWAGPR.2017.7996109

§  A.RisDć,M.Vrtunski,M.Govedarica,L.Pajewski,X.Derobert,“AutomatedDataExtrac-onfromSynthe-candRealRadargramsofDistrictHea-ngPipelines”,IWAGPR2017,9thInternaDonalWorkshop,Edinburgh,UK,June28-302017,DOI:10.1109/IWAGPR.2017.7996046

§  A.RisDć,Ž.Bugarinović,M.Vrtunski,M.Govedarica,D.Petrovački,“Integra-onofmodernremotesensingtechnologiesforfasteru-litymappinganddataextrac-on”,ConstrucDonandBuildingMaterials,Volume154,November2017,Pages1183-1198

More information

10


Comparison with SPOT-GPR The following slides show a comparative overview of the results of 1.  SPOT-GPR: a freeware tool for target detection and localization

in GPR data developed within the COST Action TU1208 2.  Automated point coordinates extraction from localized hyperbolic

reflections in GPR data (APEX)


Test-case geometric model (synthetic data created using gprMax)

12


APEX – point extraction results (slide 1/5)

13


Cell1-1a)Object Hyp.posi-onerror[m] SAP-DOAposi-onerror[m] APEXposi-onerror[m]No.1:Lefedge 0.0001 -0.008 -0.0028 0.01155 -0.0015 0.0075No.2:Lef -0.0012 -0.01 -0.0016 0.009 -0.0026 0.0118No.3:Centre 0.001 -0.011 0.00035 0.005 -0.0092 0.0164No.4:Right 0.0031 -0.005 0.0004 0.004 0.0023 0.0129No.5:Rightedge 0.007 -0.015 0.0054 0.0112 - -

Cell1-1b)Object Hyp.posi-onerror[m] SAP-DOAposi-onerror[m] APEXposi-onerror[m]No.1:Lefedge 0.0001 -0.008 -0.0028 0.01155 -0.0018 0.0075No.2:Lef -0.0012 -0.01 -0.0016 0.009 -0.0017 0.0129No.3:Centre 0.001 -0.011 0.00035 0.005 0.0007 0.0171No.4:Right 0.0031 -0.005 0.0004 0.004 0.0026 0.0129No.5:Rightedge 0.007 -0.015 0.0054 0.0112 - -

Cell1-1c)Object Hyp.posi-onerror[m] SAP-DOAposi-onerror[m] APEXposi-onerror[m]No.1:Lefedge -0.0012 -0.008 -0.0018 0.0286 -0.0017 0.0075No.2:Lef -0.003 -0.01 -0.0048 0.0184 -0.0015 0.0129No.3:Centre -0.002 -0.011 -0.007 0.005 -0.0006 0.0171No.4:Right 0.003 -0.005 0.00001 0.01336 0.0026 0.0129No.5:Rightedge 0.007 -0.015 0.001 0.0142 - -

Cell1-1d)Object Hyp.posi-onerror[m] SAP-DOAposi-onerror[m] APEXposi-onerror[m]No.1:Lefedge -0.0063 -0.008 -0.008 0.0294 -0.002 0.0075No.2:Lef -0.0015 -0.01 -0.0042 0.0184 -0.0017 0.0129No.3:Centre -0.003 -0.011 -0.005 0.0043 -0.0007 0.0171No.4:Right 0.0126 -0.0045 -0.005 0.0135 0.0022 0.0129No.5:Rightedge 0.007 -0.0155 0.008 0.014 - -

Comparative results: Cell 1-1 (slide 2/5)

14


Cell 1-1 graphical overview (slide 3/5)

15

-0,015

-0,01

-0,005

0

0,005

0,01

0,015

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16

[m]

Horizontalposi-on-Cell1-1

FiqngHyp.

SAP-DOA

APEX

-0,015

-0,01

-0,005

0

0,005

0,01

0,015

0,02

0,025

0,03

0,035

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16

[m]

Ver-calposi-on-Cell1-1

FiqngHyp.

SAP-DOA

APEX


Cell1-2a)Object Hyp.posi-onerror[m] SAP-DOAposi-onerror[m] APEXposi-onerror[m]No.1:Lefedge -4.72E-04 0.0149 -0.0035 0.0107 -0.002 0.0121No.2:Lef -0.0018 0.0054 0.001 0.011 0.0026 0.0029No.3:Centre 0.003 0.0064 -0.0036 0.0171 -0.0036 0.0039No.4:Right 2.80E-05 0.0338 0.008 0.0193 0.0023 0.0068

Cell1-2b)Object Hyp.posi-onerror[m] SAP-DOAposi-onerror[m] APEXposi-onerror[m]No.1:Lefedge -0.019 0.00147 -0.0025 -0.0126 -0.0022 0.0121No.2:Lef 0.0951 0.068 -0.001 0.017 0.0027 0.0029No.3:Centre 0.01 0.059 -0.0053 -0.0241 0.0007 0.0039No.4:Right 0.02 0.038 0.0056 0.026 0.0019 0.0068

Cell1-2c)Object Hyp.posi-onerror[m] SAP-DOAposi-onerror[m] APEXposi-onerror[m]No.1:Lefedge -0.0043 0.0148 -0.003 0.0138 -0.0024 0.0121No.2:Lef -0.0032 0.0051 -0.0003 0.0139 0.0024 0.0029No.3:Centre 0.0165 0.0043 -0.0022 0.0139 0.0009 0.0039No.4:Right -0.0015 0.0037 -0.0003 0.0241 0.0019 0.0082

Cell1-2d)Object Hyp.posi-onerror[m] SAP-DOAposi-onerror[m] APEXposi-onerror[m]No.1:Lefedge -0.0127 0.015 -0.003 0.0134 -0.0024 0.0121No.2:Lef 0.0015 0.01 -0.0007 0.0167 0.0022 0.0029No.3:Centre 0.003 0.011 -0.003 0.017 0.0002 0.0039No.4:Right -0.0126 0.0045 0.00001 0.008 0.0011 0.0082

Comparative results: Cell 1-2 (slide 3/5)

16


Cell 1-2 graphical overview (slide 4/5)

17

-0,04

-0,02

0

0,02

0,04

0,06

0,08

0,1

0,12

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16

[m]

Horizontalposi-onerror-Cell1-2

FiqngHyp.

SAP-DOA

APEX

-0,04

-0,02

0

0,02

0,04

0,06

0,08

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16

[m]

Ver-calposi-onerror-Cell1-2

FiqngHyp.

SAP-DOA

APEX


•  SAP-DOAvsAPEX

Results - summary (slide 5/5)

18

-0,006

-0,004

-0,002

0

0,002

0,004

0,006

0,008

0,01

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16

[m]

Horizontalposi-onerror-Cell1-2

SAP-DOA

APEX

-0,03

-0,02

-0,01

0

0,01

0,02

0,03

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16

[m]

Ver-calposi-onerror-Cell1-2

SAP-DOA

APEX


Summary (slide 1/2)

Algorithm1(hyperbolafiYng)§  Yieldslessaccurateresultsfordepthcoordinate(asexpected)§  The main problem of algorithm 1 is apex esDmaDon in case of

crossed hyperbolic reflecDons, and also in situaDon whenhyperbolic reflecDons are one beneath the other (the way thepointsforfiqngareselectedisproblemaDc)

§  PosiDoncoordinatesaredeterminedwithsufficientaccuracyAlgorithm2(SAP-DOA)§  Yieldsverygoodresults forbothcoordinatesoftheapex(posiDon

anddepth)§  Produces resultswhen other prong of hyperbolic reflecDon is not

visible(Cell1-1,No.5–rightedge)

19


Algorithm3(APEX)§  ResultsaresimilartoSAP-DOA§  Bestresultsindeterminingthedepth§  ApexofhyperbolicreflecDon(Cell1-1,No.5–rightedge)wasnot

detectedsinceenDrerightprongwasmissing,andtheapexwasontheveryedgeofthedomain

Summary (slide 2/2)

20


§  BothalgorithmsperformedwellonsyntheDcdata§  A similar comparison should be done on experimental data (COST

TU1208opendatabaseofrardargramscanbeuseful)§  Both algorithms can be applied on objects of non-cylindrical shape,

suchasaconcretechannelcontainingheaDngpipelines§  Extracted points from hyperbolic reflecDons can be used to esDmate

thevaluesofsomeotherparameters(e.g.uDlitydiameterandaveragevelocity)byapplyingappropriatemodelandfiqng

Conclusions

21


Thank you for your attention!

Aleksandar Ris.ć [email protected]

development of a new algorithm for automated point ... · ² segregaon of small two-dimensional...

Documents