Tracking Learning Detection A Camera that Learns

Embed Size (px)

Citation preview

Tracking Learning Detection

TrackingLearningDetectionA Camera that Learns.1Why is object tracking so important ?

surveillance

hand recognition

scene analysisvehicle detectionrobot visionmotion capture

3Outline- Framework- Key ideasWhat is TLD intuitively?How does TLD work exactly?Experiment ResultsWhy TLD?- Traditional methods- Challenges to old ways- Tracking- Learning- Detection- TLDWhy TLD ?

Why TLD ?

https://www.youtube.com/watch?v=m-uqKfzzkvM?6Why TLD ?- Traditional method 1 :TrackingAccumulate error.Fail if the object disappears. Only need initialization.Appearance changing & Partial occlusions. (?)

Estimate motion over frames

7

Why TLD ?- Traditional method 2 : DetectionOff-lineLearning

Do not fail if the object disappears. Need off-line training.Appearance changing & Partial occlusions.

Scan every frame independently8Why TLD ?- Challenges to old ways

Not fail if the object disappears. Appearance changing. Partial occlusions.9Why TLD ?- Challenges to old ways

Not fail if the object disappears. Appearance changing. Partial occlusions.

TLD is an algorithm that can solve all these tasks !10What is TLD intuitively?

What is TLD intuitively?- FrameworkTrackingDetectionOff-lineLearning12What is TLD intuitively?- FrameworkTrackingDetectionOn-lineLearningUpdateLocationTrajectoryfragmentUpdate13What is TLD intuitively?- Key ideas 1 : Learning

Track & detectTrackDetectUpdatestraining examples(add/delete)Update detector

Appearance changing. 14What is TLD intuitively?- Key ideas 2 : Updating

True trajectoryTrackingDetectionTLD

Not fail if no object.

Partial occlusions.How does TLD work exactly?

16How does TLD work exactly?- Framework

How does TLD work exactly?- Tracking

1.Reliable point selection:

Choose point 1Discard point 2- Use forward-backward selection to select 50% of the most reliable points.- Forward-backward selection

18How does TLD work exactly?- Tracking

2.Median-Flow tracker:

Tracks the reliable points between consecutive frames ,and estimates displacement using median with confidence proportional to (1/scale of the median).19How does TLD work exactly?- Framework

How does TLD work exactly?- Detection

2bitBPfeaturesScanningwindow

2bitBP featuresIlluminationinvariance2BitBP : invariance to illuminationRF : fast21How does TLD work exactly?- Detection

RandomForestPosteriors2bitBPfeaturesScanningwindow

Random ForestIf the average posteriors is larger than 50%, consider it an object

2BitBP : invariance to illuminationRF : fast22How does TLD work exactly?- Detection

1-NN classifierConfidence =d-/(d-+d+)RandomForestPosteriors2bitBPfeaturesScanningwindow

Feature space

23How does TLD work exactly?- Framework

Pick the one with higher confidence scoreHow does TLD work exactly?- Learning

Feature space

25How does TLD work exactly?- Learning

Track & detectTrackDetectUpdatestraining examples(add/delete)Update detector

26How does TLD work exactly?- Learning :

1610Appearance changingproblem solved!9

27How does TLD work exactly? - Recap : t=0

How does TLD work exactly? - Recap : t=1

Decide by confidence score

How does TLD work exactly? - Recap : t=1

How does TLD work exactly? - Recap : t=2

Remove if false detect

Experiment ResultExperiment Result

More vid33

Experiment Result

Experiment ResultThank you