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

COSTissupportedbytheEUFrameworkProgrammeHorizon2020 TheFinalConferenceoftheCOSTAcDonTU1208,Warsaw,2017

General steps of the automated algorithm

5.Dataextrac-on

4.Detectionandremovalofinterferedhyperbolas

2.Findingapexes 3.Findingpointsontheprongs

1.Detec-onofsegmentsofinterest(SOI)

TheFinalConferenceoftheCOSTAcDonTU1208,Warsaw,2017

§  STEP1:Conversiontorasterformatusingrad2bmp.§  STEP2:LocaDngofSOIusingMatLABGUItrainCascadeObjectDetector

(COD)Ø  Trainingset:posiDveandnegaDvetrainingsamplesØ  Usingrealand/orsyntheDcdata

1. Detection of segments of interest (SOI)

4

TheFinalConferenceoftheCOSTAcDonTU1208,Warsaw,2017

2. Finding apexes (slide 1/2)

§  STEP1:FindingtheiniDalpixelforsearchinSOI§  STEP2:Formingasub-matrixwithdimensions3x2aroundofiniDalpixel

withcurrentpixelinthesecondrowandthefirstcolumn§  STEP3:Findinglocalmaximumsinsub-matrix3x2(iteraDvely)

5

TheFinalConferenceoftheCOSTAcDonTU1208,Warsaw,2017

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

TheFinalConferenceoftheCOSTAcDonTU1208,Warsaw,2017

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

TheFinalConferenceoftheCOSTAcDonTU1208,Warsaw,2017

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

TheFinalConferenceoftheCOSTAcDonTU1208,Warsaw,2017

§  WithminormodificaDons,thisalgorithmcanalsobeappliedtothecaseofdistrictheaDngsystems

Algorithm applications on complex structures

9

TheFinalConferenceoftheCOSTAcDonTU1208,Warsaw,2017

§  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

COSTissupportedbytheEUFrameworkProgrammeHorizon2020 TheFinalConferenceoftheCOSTAcDonTU1208,Warsaw,2017

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)

TheFinalConferenceoftheCOSTAcDonTU1208,Warsaw,2017

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

12

TheFinalConferenceoftheCOSTAcDonTU1208,Warsaw,2017

APEX – point extraction results (slide 1/5)

13

TheFinalConferenceoftheCOSTAcDonTU1208,Warsaw,2017

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

TheFinalConferenceoftheCOSTAcDonTU1208,Warsaw,2017

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

TheFinalConferenceoftheCOSTAcDonTU1208,Warsaw,2017

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

TheFinalConferenceoftheCOSTAcDonTU1208,Warsaw,2017

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

TheFinalConferenceoftheCOSTAcDonTU1208,Warsaw,2017

•  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

TheFinalConferenceoftheCOSTAcDonTU1208,Warsaw,2017

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

TheFinalConferenceoftheCOSTAcDonTU1208,Warsaw,2017

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

detectedsinceenDrerightprongwasmissing,andtheapexwasontheveryedgeofthedomain

Summary (slide 2/2)

20

TheFinalConferenceoftheCOSTAcDonTU1208,Warsaw,2017

§  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

COSTissupportedbytheEUFrameworkProgrammeHorizon2020 TheFinalConferenceoftheCOSTAcDonTU1208,Warsaw,2017

Thank you for your attention!

Aleksandar Ris.ć aris.c@uns.ac.rs