360 likes | 471 Vues
T racking L earning D etection. A Camera that Learns. Why is object tracking so important ?. vehicle detection. motion capture. r obot vision. scene analysis. hand recognition. surveillance. Outline. Why TLD?. - Traditional methods - Challenges to old ways. What is TLD intuitively?.
E N D
TrackingLearningDetection A Camera that Learns.
Why is object tracking so important ? vehicle detection motion capture robot vision scene analysis hand recognition surveillance
Outline Why TLD? - Traditional methods - Challenges to old ways What is TLD intuitively? - Framework - Key ideas Howdoes TLD work exactly? - Tracking - Learning - Detection - TLD Experiment Results
Why TLD ?- Traditional method1 : Tracking Estimate motion over frames Only need initialization. Appearance changing & Partial occlusions.(?) Accumulate error. Fail if the object disappears.
Why TLD ?- Traditional method2 : Scan every frame independently Off-line Learning Detection Do not fail if the object disappears. Need off-line training. Appearance changing & Partial occlusions.
Why TLD ?- Challenges to old ways • Not fail if the object disappears. • Appearance changing. • Partial occlusions.
Why TLD ?- Challenges to old ways • Not fail if the object disappears. TLD is an algorithm that can solve all these tasks ! • Appearance changing. • Partial occlusions.
What is TLD intuitively?- Framework Tracking Off-line Learning Detection
What is TLD intuitively?- Framework On-line Learning Location Trajectory fragment Update Tracking Detection Update
What is TLD intuitively?- Key ideas1 : Learning Updates training examples (add/delete) Detect Track & detect Update detector • Appearance changing. Track
What is TLD intuitively?- Key ideas2 : Updating True trajectory Tracking Detection TLD • Not fail if no object. • Partial occlusions.
How does TLD work exactly?-Tracking 1.Reliable point selection: - Forward-backward selection Choose point 1 Discard point 2 - Use forward-backward selection to select 50% of the most reliable points.
How does TLD work exactly?-Tracking 2.Median-Flow tracker: and estimates displacement using median with confidence proportional to (1/scale of the median). Tracks the reliable points between consecutive frames ,
How does TLD work exactly?-Detection 2bitBP features Scanning window Illumination invariance 2bitBP features
How does TLD work exactly?-Detection Random Forest Random Forest Posteriors Scanning window If the average posteriors is larger than 50%, consider it an object 2bitBP features
How does TLD work exactly?-Detection 1-NN classifier Confidence =d-/(d-+d+) Feature space Random Forest Posteriors Scanning window 2bitBP features
How does TLD work exactly?-Framework Pick the one with higher confidence score
How does TLD work exactly?-Learning Feature space
How does TLD work exactly?-Learning Updates training examples (add/delete) Detect Track & detect Update detector Track
How does TLD work exactly?-Learning : Appearance changing problem solved! 1 6 10 9
How does TLD work exactly? - Recap : t=1 Decide by confidence score
How does TLD work exactly? - Recap : t=2 Remove if false detect