Reconstruction of a road by local image matches and global 3D optimization
ABSTRACT A method is presented for reconstructing a 3-D road from a single
image. It finds the images of opposite points of the road. Opposite
points are points which face each other on the opposite sides of the
road; the images of these points are called matching points. For points
chosen from one side of the road image, the algorithm finds all the
matching point candidates on the other side, based on local properties
of a road. However, these solutions do not necessarily satisfy the
global properties of a typical road. A dynamic programming algorithm is
applied to reject the candidates which do not fit the global road. A
benchmark using synthetic roads is described. It shows that the roads
reconstructed by the proposed method match the actual roads better than
those reconstructed by two other road reconstruction algorithms.
Experiments with 50 road images taken by the autonomous land vehicle
(ALV) showed that the method is robust with real-world data and that the
reconstructions are fairly consistent with road profiles obtained by
fusion between range images and video images
- SourceAvailable from: Dieter Koller[Show abstract] [Hide abstract]
ABSTRACT: We propose a new approach for vision based longitudinal and lateral ve- hicle control which makes extensive use of binocular stereopsis. Longitudi- nal control - i.e. maintaining a safe, constant distance from the vehicle in front - is supported by detecting and measuring the distances to leading vehicles using binocular stereo. A known camera geometry with respect to the locally planar road is used to map the images of the road plane in the two camera views into alignment. Any significant residual image disparity then indicates an object not lying in the road plane and hence a potential obstacle. This approach allows us to separate image features into those lying in the road plane, e.g. lane markers, and those due to other objects. The features which lie on the road are stationary in the scene and appear to move only because of the egomotion of the vehicle. Measurements on these features are used for dynamic update of (a) the camera parameters in the presence of camera vibration and changes in road slope (b) the lateral position of the vehicle with respect to the lane markers. In the absence of this separation, image features due to vehicles which happen to lie in the search zone for lane markers would corrupt the estimation of the road boundary contours. This problem has not yet been addressed by any lane marker based vehicle guid- ance approach, but has to be taken very seriously, since usually one has to cope with crowded traffic scenes where lane markers are often obstructed by vehicles. Lane markers are detected and used for lateral control, i.e. following the road while maintaining a constant lateral distance to the road boundary. For that purpose we model the road and hence the shape of the lane markers as clothoidal curves, the curvatures of which we estimate recursively along the image sequence. These curvature estimates also provides desirable look-ahead information for a smooth ride in the car.
- [Show abstract] [Hide abstract]
ABSTRACT: This paper presents the visual servoing of a six degrees of freedom (6-DOF) manipulator for unknown three-dimensional profile following. The profile has an unknown curvature, but its cross section is known. The visual servoing keeps the transformation between a cross section of the profile and the camera constant with respect to 6 DOE The position of the profile with respect to only five degrees of freedom can be measured with the camera since the image does not provide position information along the profile. The kinematic model of the robot is used to reconstruct the displacement along the profile, i.e., the sixth degree of freedom, and allows to control the profile-following velocity. Experiments show good accuracy for positioning at a sampling rate of 50 Hz. Two control strategies are tested: proportional-integral control and generalized predictive control (GPC). The visual servoing exhibits better accuracy with the GPC in simulations and in real experiments on a 6-DOF manipulator due to the predictive property of the algorithm.IEEE Transactions on Robotics and Automation 08/2002; 18(4):511-520.
- [Show abstract] [Hide abstract]
ABSTRACT: 4D/RCS is a hierarchical architecture designed for the control of intelligent systems. One of the main areas that 4D/RCS has been applied to recently is the control of autonomous vehicles. To accomplish this, a hierarchical decomposition of on-road driving activities has been performed which has resulted in implementation of 4D/RCS tailored towards this application. This implementation has seven layers and ranges from a journey manager which determines the order of the places you wish to drive to, through a destination manager which provides turn-by-turn directions on how to get to a destination, through a route segment, drive behavior, elemental maneuver, goal path trajectory, and then finally to servo controllers. In this paper, we show, within the 4D/RCS architecture, how knowledge-driven top-down symbolic representations combined with low-level bottom-up tasks can synergistically provide valuable information for on-road driving better than what is possible with either of them alone. We demonstrate these ideas using field data obtained from an Unmanned Ground Vehicle (UGV) traversing urban on-road environments.Proceedings of SPIE - The International Society for Optical Engineering 05/2007;
Anew methodis presentedforreconstructinga ?Droadfromasingleimage? It
?nds theimagesofoppositepoin ts oftheroad?oppositepointsarepoints whichface
eachotheronthe opposite sidesof the road?theimagesof thesepointsarecalled
matching points??Forpoin tschosen fromoneside oftheroadimage?the prop osed
algorithm?ndsallthe matc hingpoin tcandidatesonthe other side?based onlocal
propertiesofaroad?Ho w everthese solutionsdonot necessarilysatisfy the global
propertiesofatypicalroad?Adynamic programmingalgorithm isappliedtoreject
the candidateswhichdo not?t the globalroad?
Abenchmark usingsynthetic roadsisdescribed?whichshows thatthe roads re?
constructedby theprop osedmethodmatch theactualroadsbetterthantwoother
roadreconstruction algorithms? Experiments with??road imagestakenby theAu?
tonomousLandVehicle?ALshothatdisrobustwith realw orld data?
andthatthe reconstructionsarefairly consistent withroadpro?lesobtained byfusion
roads?????? ???? ???Forrobustnessinavarietyofconditions?these systemscan be
driv enbya sup ervisor systemreasoning ab outinformation pro videdbyseveralalgorithms?
suchasstereoalgorithms?stereomotionalgorithms?algorithms usingsinglevideo frames
or combiningvideo framesandrangeimages? Kahlmanpredictorscombining information
obtainedfrom several vehiclepositions?etc????Some algorithmsmay monitortheroadover
ashort distance oralongasingleedge?forinput toafaststeering controlloop ????Other
algorithmsmay attempttoextendtheir analysis tothe mostdistantavailabledata infront
ofthevehicle? forinputtolonger termreasoning mo dules?
Thispap erpresents a newalgorithm abletoreconstructtheroadshape froma single
image? providing thethreedimensional pro?leofthe roadinfrontofthe vehicle?often
distance presentsseveraladvantages?Thereconstructionsfromseveral videoframes can
beov erlapped?andtheevidencefrom each reconstructioncan becombined for added
reliability?It isalsoclearly usefulforaroadfollowing systemtomakeestimationsof turns
wellinadvance?andadjust itssp eedaccordingly?Thislong rangeobservationoftheroad
doesnot precludethe useofashortrangeroadanalysisinthecontrolloopof the vehicle
Roadreconstructionfromasingleimageisa ?shape?from?contour?problem? Itis ob?
viouslyunder?constrained?yieldinganin?nit yofpossibleroadshapes unlessconstrain ts
abouttheroad structureinthe?Dscenearein troduced? Thus aroad modelhastobe
assumed?whic hprovidesareasonable set ofadditional constraints?
The simplestmo delwhichhasbeenapplied ????is theFlat?Earthge ometry model?the
roadis assumedplanarand inthesameplanewhichsupportsthevehicle?andtheroad
imageanalysisgivesraggedroadimageedges?Butitisvery sensitiveto thedi?erence
betweentheassumedand actualcameratilt anglewithrespectto the ground?Figure ???
F oracamera mounted onav ehicleat ???metersab ovethe ground?aw orldp ointat??
metersin frontof thev ehiclewillbe reconstructedat??meters ifthegroundplaneangle
The constantroad widthconstraintisnotsu?cient?Another constraintmustbeadded
for thereconstructiontobepossible? Weha vec hosenthe zer o?bankconstraint? specifying
that theroaddo es nottilt sideways? Aroad modelcombiningconstantwidth andzero
bank wasoriginally suggestedin?????
In previouswork?wedevelopedanincrementalroadreconstructionmethod basedon
theseconstraints ?????inwhichanewpairofedgep oints couldbe found ifwehadalready
edge poin tsclosetothe vehicle toedge p ointsin thedistance?Thismethod isfragile
because anyincrementof constructiondep endsonthe previouselementsinthechain?Any
failureof theroad reconstructionatanypoin tcan befataltothefurtherprogressof the
Thisincrementalmetho dusedadiscreteapproach?Roadreconstructions basedona
di?eren tial approachcanbefoundin??? ??? Aninteresting alternativetothe globaldynamic
programmingoptimizationproposed inthepresen tpapercan befound in????
Theprop osedalgorithm canbedecomp osedintothefollowingsteps?
??Inapreliminarystep? notdetailedhere?appropriateimagepro cessing techniqueshave
isolated thetwo curvesof the edgesinthe image?andapolygonalapproximationhas
beenfound foreach edgecurv e?
imagepoin tsarecalled matchingpoints if theyare imagesof theendpoints ofcross?
segments ofthe ?Droad??This matc hing is madepossible by makingreasonable
hyp otheseson theshape of theroad? whichaddenoughconstraintsto makethe
problem solvable?Sp eci?cally?theroad ismodelledasaspaceribbonde?nedbya
centerline spineandhorizontalcross?segmentsofconstantlength cuttingthe spineat
theirmidpointsatanormaltothespine?Wefurtherassume thattangen tstothe
Vis thevertical direction?F oredge curv es approximatedbypolygonal lines?
canbeonaline segment?anditsposition betweentheend
points ofthelinesegmentcan beexpressedbyanumberbetween?and?? whereas
ofalinesegmen t?withaconstantp ositionbut witha tangentanglewhichcan be
expressedbyan umber between?and? withintherangeofanglesofthetwoadjacent
linesegmen ts ?Section ???F oreac hpointchosen fromone imageedge?wecheckfor
eachofthe linesegmentsofthe otherimage edgeif a matchingpoin tb elongs tothat
line segment?i?e?? ifourexpressiongives alinearcoordinatebet w een?and?forthis
line segment?Then welook for matchingpointsatthe nodesof thepolygonalline
byc heckingiftheexpressiongivesanumberbetween? and? forthe tangen tangle?
canbevery roughandwiggly? Anotherreason isthatthecondition usedisonly a
necessary conditionfor twop oints tobe matchingpointsintheimageof theroad?
This conditionislo cal andwemuststill c ho osethematc hingpoin tspairswhichare
themostgloballyconsisten t? and discardtheotherpairs?Thecriteriaofoptimization
befound?This correspondence isunique ifthe cross?segmentsareassumedhorizontal
whichcharacterizesa?goodroad?? Thetotalev aluationfunctionis thesum of the
functionsofeachof thearcs ofthe graph?The evaluationfunctionfor anarc isthe
sumofweigh tedcriteria?which gradethec hoices ofindividualcross?segmen tsand
theneighb orhoodofconsecutivecross?segments? basedonangular considerations?
?Thematchingpoin t problem
Considerthe imageofarailroadtrackand itsrailroadties?and assumethatsomeappro?
priate image processing techniqueshavereducedthe images ofthe rails tocurvesand the
images of theties to linesegmentsb etweenthesecurves?Figure ???Thepositions ofthe
end points ofthe tiesegmentsonthe curvesofthe railarethe matchingpointsinthe
image?Thereconstruction oftheshape oftherailroadtrackin?Dspaceuses thematching
pointsandisstraigh tforwardifthreeh yp othesesare made?
??The widthwoftherailroadtrac kis constantand known?
??Thecoordinates of theverticalunitvector
V arekno wn in thecamera coordinate
??Therailroad tiesare approximately horizon tal?
Obviously? the lasth ypothesisdoes notconstrain therailroaditself tobe horizon tal?Simi?
?theendp oints ofthe image ofatie?The
correspondingvectors from theviewp ointO to theseimagep ointswillbe denotedby?a
? The correspondingw orldp oints A
sinceworldpoin ts and theirimages are on thesameline ofsigh t?
Thew orldlinesegment isassumedhorizontal? thetwoparameters?