Transcript
Page 1: CSTalks - Object detection and tracking - 25th May

1

Tutorial 7Object Detection and Tracking

Mukesh Saini

Page 2: CSTalks - Object detection and tracking - 25th May

2

LayoutObject detection

ChallengesGMM based method

TrackingChallengesParticle filter based tracking

Page 3: CSTalks - Object detection and tracking - 25th May

3

Object DetectionGoal: To detect the regions of the image that are

semantically important to us:PeopleVehicleBuildings

ApplicationCrowd managementTraffic managementVideo compression, video surveillance, vision-

based control, human-computer interfaces, medical imaging, augmented reality, and robotics…

Page 4: CSTalks - Object detection and tracking - 25th May

4

Object Detection in ImagesSubjectively definedGenerally template basedMainly done by image segmentation

Page 5: CSTalks - Object detection and tracking - 25th May

5

Object Detection in VideosRelatively Moving – ObjectRelatively Static - Background

The goal here is to differentiate the moving object from background!

Page 6: CSTalks - Object detection and tracking - 25th May

6

Surveillance VideoStatic cameraBackground relatively staticSubtract the background image from current

imageIdeally this will leave the moving objectsThis is not an ideal world…

Page 7: CSTalks - Object detection and tracking - 25th May

7

Feature BasedObjects are modeled in terms of featuresFeatures are chosen to  handle changes in

illumination, size and orientationShape based – Very hard Color based – Low cost but not accurate

Page 8: CSTalks - Object detection and tracking - 25th May

8

Template BasedExample template are givenObject detection becomes matching featuresImage subtraction, correlation

Page 9: CSTalks - Object detection and tracking - 25th May

9

Motion BasedModel backgroundSubtract from the current imageLeft are moving objects Remember! This is not a real world…

Page 10: CSTalks - Object detection and tracking - 25th May

10

Problems in Modeling BackgroundAcquisition noiseIllumination variationClutterNew object introduced into backgroundObject may not move continuously

Page 11: CSTalks - Object detection and tracking - 25th May

11

Outline of Object DetectionDetermine the background and foreground

pixelsDraw contours around foreground pixelsUse heuristics to merge these contours

Page 12: CSTalks - Object detection and tracking - 25th May

12

Ideal WorldSingle value modeling of backgroundAnything different is foreground

Page 13: CSTalks - Object detection and tracking - 25th May

13

Static BackgroundEach pixel resulted from a particular surface

under particular lighteningSingle Gaussian is enough ( )

IfPixel belongs to background, else foreground

,

*5.2tP

Background ForegroundForeground

Page 14: CSTalks - Object detection and tracking - 25th May

14

Whenever a pixel matches the background Gaussian, update the background model i.e.IfThenStandard deviation updated accordingly

Illumination Variation

tP tt *5.2ttt )1(1

Page 15: CSTalks - Object detection and tracking - 25th May

15

ClutterThink of tree leaves…Multiple surfaces, still part of backgroundGaussian Mixture ModelUpdate each Gaussian after matching

Page 16: CSTalks - Object detection and tracking - 25th May

16

Static Object IntroducedThink of flower pot…Background model should adapt to this

changeUse Gaussian for new surface as well

Few extra Gaussians for the foreground

Page 17: CSTalks - Object detection and tracking - 25th May

17

Measuring PersistenceModeled as prior weight wMore persistent Gaussians belong to

backgroundIf a new pixel does not match to any exiting

Gaussians, least persistent Gaussian is replaced with a new Gaussian with:

And standard variation = a large value

tt P

t

Page 18: CSTalks - Object detection and tracking - 25th May

18

Background SelectionA background Gaussian will have

More persistence – high wLess variation – low Sort Gaussians wrt

Pick top k Gaussians as background such that

If pixel belongs to one of these, it’s a background pixel

Twk

iik

1

minarg

ttw /

Page 19: CSTalks - Object detection and tracking - 25th May

19

Adaptive Background ModelEvery pixel is modeled as mixture of

GaussiansMore persistent Gaussians belong to

background and others to foregroundThe Gaussians are updated after each frame

Page 20: CSTalks - Object detection and tracking - 25th May

20

Connecting the DotsThe output of background modeling is a

binary imageDilation/Erosion can further reduce noiseContour drawingBounding boxes

Page 21: CSTalks - Object detection and tracking - 25th May

21

Revisit the problemsProblems

Slow moving background – clutterNew object introduced into backgroundIllumination variationObject may not move continuously

Page 22: CSTalks - Object detection and tracking - 25th May

22

Thank YouQ & A


Top Related