development of a new algorithm for automated point ... · ² segregaon of small two-dimensional...
TRANSCRIPT
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 [email protected]