11.07.2015 Views

Download Presentation in PDF (2.26 MB)

Download Presentation in PDF (2.26 MB)

Download Presentation in PDF (2.26 MB)

SHOW MORE
SHOW LESS

Create successful ePaper yourself

Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.

Why hasn’t it been solved yet?Big cities pose anextremely hardproblem because ofmultiple light sources.Bad weather makesdetection even moredifficult.Night driv<strong>in</strong>g is alsoa challeng<strong>in</strong>gscenario.Comb<strong>in</strong><strong>in</strong>g all of theabove, even <strong>in</strong> smallscale, makesdetection almostimpossible.


What has been tried?• Color threshold<strong>in</strong>g <strong>in</strong> RGB or HSV• Symmetry Detection with Hough or othernovel methods.• Structural Models of Traffic Lights.• Track<strong>in</strong>g modules to help m<strong>in</strong>imize falsepositive results.• A mixture of the above.


Image Pre-Process<strong>in</strong>gStep 4: Bloom<strong>in</strong>g Effect ReductionR-G:Y-B:Image fill<strong>in</strong>g of the twoimages followed by addition.This reduces the “bloom<strong>in</strong>geffect” significantly.


Image Pre-Process<strong>in</strong>gStep 4: Bloom<strong>in</strong>g Effect ReductionR-G:Y-B:Image fill<strong>in</strong>g of the twoimages followed by addition.This reduces the “bloom<strong>in</strong>geffect” significantly.


Image Pre-Process<strong>in</strong>gStep 1: RGB to L*a*b*L*:a*:b*:


Image Pre-Process<strong>in</strong>gStep 2: Red – Green EnhancementL*:a*:Multiply<strong>in</strong>g L* and a* toenhance only bright redand green regions(exclude trees, roofs, etc.)


Image Pre-Process<strong>in</strong>gStep 3: Yellow – Blue EnhancementL*:Multiply<strong>in</strong>g L* and b* toenhance only bright yellow andblue regions(red lights <strong>in</strong>clude yellow andgreen lights <strong>in</strong>clude some blue)b*:


Image Pre-Process<strong>in</strong>gStep 4: Bloom<strong>in</strong>g Effect ReductionR-G:Y-B:Image fill<strong>in</strong>g of the twoimages followed by addition.This reduces the “bloom<strong>in</strong>geffect” significantly.


Image Pre-Process<strong>in</strong>gStep 4: Bloom<strong>in</strong>g Effect ReductionR-G:Y-B:Image fill<strong>in</strong>g of the twoimages followed by addition.This reduces the “bloom<strong>in</strong>geffect” significantly.


Traffic Light Candidate Detection


Traffic Light Candidate DetectionStep 1: FastRadial Transform


Traffic Light Candidate DetectionStep 1: FastRadial Transform


Traffic Light Candidate DetectionStep 2:Maxima/M<strong>in</strong>imaLocalization


Traffic Light Candidate Detection


Traffic Light Candidate DetectionStep 1: FastRadial Transform


Traffic Light Candidate DetectionStep 1: FastRadial Transform


Traffic Light Candidate DetectionStep 2:Maxima/M<strong>in</strong>imaLocalization


Traffic Light Candidate VerificationFRAME #1:Spatiotemporalpersistency check:Traffic Lightscandidates for 4consecutive framesget verified.11 11111


Traffic Light Candidate VerificationFRAME #2:Spatiotemporalpersistency check:Traffic Lightscandidates for 4consecutive framesget verified.22 221


Traffic Light Candidate VerificationFRAME #3:Spatiotemporalpersistency check:Traffic Lightscandidates for 4consecutive framesget verified.132113 1


Traffic Light Candidate VerificationFRAME #4:Spatiotemporalpersistency check:Traffic Lightscandidates for 4consecutive framesget verified.2112142


Traffic Light Candidate VerificationFRAME #5:Spatiotemporalpersistency check:Traffic Lightscandidates for 4consecutive framesget verified.2 11235


Traffic Light Candidate VerificationFRAME #6:Spatiotemporalpersistency check:Traffic Lightscandidates for 4consecutive framesget verified.13223461


Experimental EvaluationQuantitative• Normal weather conditions.• City Driv<strong>in</strong>g.• Daytime.• Manually Annotated.• 8-bit, 640x480 @ 25fps• 11179 framesQualitative• Ra<strong>in</strong>y Weather• Night Driv<strong>in</strong>g• Different cameras.• Different resolutions.• <strong>Download</strong>ed off YouTube,OR shot <strong>in</strong> Greek roads.The quantitative data is com<strong>in</strong>g fromthe Robotics Centre of M<strong>in</strong>es ParisTech and ispublicly available at:http://www.lara.prd.fr/benchmarks/trafficlightsrecognition


Results <strong>in</strong> Normal Conditions


Results <strong>in</strong> Normal Conditions


Results <strong>in</strong> Ra<strong>in</strong>y Conditions


Results <strong>in</strong> Night Driv<strong>in</strong>g Scenes


Method LimitationsFalse PositivesNight <strong>in</strong> Big Cities


Conclusions• Proposed method shows good results <strong>in</strong>both normal and adverse conditions.• Many false alarms, that could be reduced bysome k<strong>in</strong>d of structural model.• Extremely difficult cases cannot be tackled.• Future work on construction of an annotateddatabase <strong>in</strong> adverse conditions.• Also, the <strong>in</strong>corporation of a track<strong>in</strong>g moduleand a color consistency module.


Bibliography


Thank you for your attention.Questions?

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!