![Page 1: “Coupled Detection and Trajectory Estimation for ...daphna/course/student lectures/dmitri hanukaev.pdf · B. Leibe, N. Cornelis, K. Cornelis, and L. Van Gool. Dynamic 3d scene analysis](https://reader033.vdocuments.net/reader033/viewer/2022042000/5e6d6f9b2ec42c69fd334f30/html5/thumbnails/1.jpg)
“Coupled Detection and Trajectory “Coupled Detection and Trajectory Estimation for MultiEstimation for Multi--Object Tracking”Object Tracking”
By B. Leibe, K. Schindler, L. Van GoolBy B. Leibe, K. Schindler, L. Van Gool
Visual object recognition course 67777Visual object recognition course 67777
Presented By: Hanukaev Dmitri Presented By: Hanukaev Dmitri Lecturer: Prof. Daphna WienshallLecturer: Prof. Daphna Wienshall
![Page 2: “Coupled Detection and Trajectory Estimation for ...daphna/course/student lectures/dmitri hanukaev.pdf · B. Leibe, N. Cornelis, K. Cornelis, and L. Van Gool. Dynamic 3d scene analysis](https://reader033.vdocuments.net/reader033/viewer/2022042000/5e6d6f9b2ec42c69fd334f30/html5/thumbnails/2.jpg)
The GoalThe Goal
The goal is multiply object tracking by detection with application on pedestrians.
Visual object recognition course 67777Visual object recognition course 67777
![Page 3: “Coupled Detection and Trajectory Estimation for ...daphna/course/student lectures/dmitri hanukaev.pdf · B. Leibe, N. Cornelis, K. Cornelis, and L. Van Gool. Dynamic 3d scene analysis](https://reader033.vdocuments.net/reader033/viewer/2022042000/5e6d6f9b2ec42c69fd334f30/html5/thumbnails/3.jpg)
ApproachApproach
Novel approach for multi-object tracking from monocular camera source, which considers object detection and space-time trajectory estimation as a coupled optimization problem.
Visual object recognition course 67777Visual object recognition course 67777
![Page 4: “Coupled Detection and Trajectory Estimation for ...daphna/course/student lectures/dmitri hanukaev.pdf · B. Leibe, N. Cornelis, K. Cornelis, and L. Van Gool. Dynamic 3d scene analysis](https://reader033.vdocuments.net/reader033/viewer/2022042000/5e6d6f9b2ec42c69fd334f30/html5/thumbnails/4.jpg)
MotivationMotivation
� Improve robustness by coupling object detection and tracking� Enhanced object model + feedback from
trajectory estimation to detection
Visual object recognition course 67777Visual object recognition course 67777
trajectory estimation to detection
� Global optimization to resolve trajectory interactions� Incorporate real-world physical constraints
![Page 5: “Coupled Detection and Trajectory Estimation for ...daphna/course/student lectures/dmitri hanukaev.pdf · B. Leibe, N. Cornelis, K. Cornelis, and L. Van Gool. Dynamic 3d scene analysis](https://reader033.vdocuments.net/reader033/viewer/2022042000/5e6d6f9b2ec42c69fd334f30/html5/thumbnails/5.jpg)
Experimental ResultsExperimental Results
Visual object recognition course 67777Visual object recognition course 67777
![Page 6: “Coupled Detection and Trajectory Estimation for ...daphna/course/student lectures/dmitri hanukaev.pdf · B. Leibe, N. Cornelis, K. Cornelis, and L. Van Gool. Dynamic 3d scene analysis](https://reader033.vdocuments.net/reader033/viewer/2022042000/5e6d6f9b2ec42c69fd334f30/html5/thumbnails/6.jpg)
�� IntroductionIntroduction�� Related WorksRelated Works�� Previous WorksPrevious Works
�� Pedestrian DetectionPedestrian Detection
Visual object recognition course 67777Visual object recognition course 67777
�� SpaceSpace--Time Trajectory EstimationTime Trajectory Estimation
�� Coupling between Detection and Coupling between Detection and Trajectory EstimationTrajectory Estimation
�� ResultsResults
![Page 7: “Coupled Detection and Trajectory Estimation for ...daphna/course/student lectures/dmitri hanukaev.pdf · B. Leibe, N. Cornelis, K. Cornelis, and L. Van Gool. Dynamic 3d scene analysis](https://reader033.vdocuments.net/reader033/viewer/2022042000/5e6d6f9b2ec42c69fd334f30/html5/thumbnails/7.jpg)
�� IntroductionIntroduction�� Related WorksRelated Works�� Previous WorksPrevious Works
�� Pedestrian DetectionPedestrian Detection
Visual object recognition course 67777Visual object recognition course 67777
�� SpaceSpace--Time Trajectory EstimationTime Trajectory Estimation
�� Coupling between Detection and Coupling between Detection and Trajectory EstimationTrajectory Estimation
�� ResultsResults
![Page 8: “Coupled Detection and Trajectory Estimation for ...daphna/course/student lectures/dmitri hanukaev.pdf · B. Leibe, N. Cornelis, K. Cornelis, and L. Van Gool. Dynamic 3d scene analysis](https://reader033.vdocuments.net/reader033/viewer/2022042000/5e6d6f9b2ec42c69fd334f30/html5/thumbnails/8.jpg)
TrackingTracking
� Trajectory initialization� Background subtraction� Detection� Divine Intervention
Visual object recognition course 67777Visual object recognition course 67777
� Target following � Mean – Shift tracking� Extended Kalman Filters� Particle Filters
![Page 9: “Coupled Detection and Trajectory Estimation for ...daphna/course/student lectures/dmitri hanukaev.pdf · B. Leibe, N. Cornelis, K. Cornelis, and L. Van Gool. Dynamic 3d scene analysis](https://reader033.vdocuments.net/reader033/viewer/2022042000/5e6d6f9b2ec42c69fd334f30/html5/thumbnails/9.jpg)
Common Challenges Common Challenges
� View point variation
� Illumination
Visual object recognition course 67777Visual object recognition course 67777
� Scale
� Deformation
� Occlusion
Image adopted from Li Fei Fei
![Page 10: “Coupled Detection and Trajectory Estimation for ...daphna/course/student lectures/dmitri hanukaev.pdf · B. Leibe, N. Cornelis, K. Cornelis, and L. Van Gool. Dynamic 3d scene analysis](https://reader033.vdocuments.net/reader033/viewer/2022042000/5e6d6f9b2ec42c69fd334f30/html5/thumbnails/10.jpg)
Intro of Detection into TrackingIntro of Detection into Tracking
A detector can be used for:� To initialize targets or to re-initialize them in case of
failure of tracking.� The output of detector can be used directly as data
source for tracking.
Visual object recognition course 67777Visual object recognition course 67777
source for tracking.
![Page 11: “Coupled Detection and Trajectory Estimation for ...daphna/course/student lectures/dmitri hanukaev.pdf · B. Leibe, N. Cornelis, K. Cornelis, and L. Van Gool. Dynamic 3d scene analysis](https://reader033.vdocuments.net/reader033/viewer/2022042000/5e6d6f9b2ec42c69fd334f30/html5/thumbnails/11.jpg)
�� IntroductionIntroduction�� Related WorksRelated Works�� Previous WorksPrevious Works
�� Pedestrian DetectionPedestrian Detection
Visual object recognition course Visual object recognition course 6777767777
�� SpaceSpace--Time Trajectory EstimationTime Trajectory Estimation
�� Coupling between Detection and Coupling between Detection and Trajectory EstimationTrajectory Estimation
�� ResultsResults
![Page 12: “Coupled Detection and Trajectory Estimation for ...daphna/course/student lectures/dmitri hanukaev.pdf · B. Leibe, N. Cornelis, K. Cornelis, and L. Van Gool. Dynamic 3d scene analysis](https://reader033.vdocuments.net/reader033/viewer/2022042000/5e6d6f9b2ec42c69fd334f30/html5/thumbnails/12.jpg)
Related worksRelated works
� K. Okuma, A. Taleghani, N. de Freitas, J. Little, and D. Lowe. A boosted particle filter: Multi-target detection and tracking. In ECCV’04.
� Detection by Adaboost classifier.� Multi-target tracking by
Visual object recognition course 67777Visual object recognition course 67777
� Multi-target tracking by Mixture Particle Filters(variation of particle filter).
![Page 13: “Coupled Detection and Trajectory Estimation for ...daphna/course/student lectures/dmitri hanukaev.pdf · B. Leibe, N. Cornelis, K. Cornelis, and L. Van Gool. Dynamic 3d scene analysis](https://reader033.vdocuments.net/reader033/viewer/2022042000/5e6d6f9b2ec42c69fd334f30/html5/thumbnails/13.jpg)
�� IntroductionIntroduction�� Related WorksRelated Works�� Previous WorksPrevious Works
�� Pedestrian DetectionPedestrian Detection
Visual object recognition course Visual object recognition course 6777767777
�� SpaceSpace--Time Trajectory EstimationTime Trajectory Estimation
�� Coupling between Detection and Coupling between Detection and Trajectory EstimationTrajectory Estimation
�� ResultsResults
![Page 14: “Coupled Detection and Trajectory Estimation for ...daphna/course/student lectures/dmitri hanukaev.pdf · B. Leibe, N. Cornelis, K. Cornelis, and L. Van Gool. Dynamic 3d scene analysis](https://reader033.vdocuments.net/reader033/viewer/2022042000/5e6d6f9b2ec42c69fd334f30/html5/thumbnails/14.jpg)
Previous WorksPrevious Works
� Pedestrian detection -B. Leibe, E. Seemann, and B. Schiele. Pedestrian detection in crowded scenes. In CVPR’05, 2005.B. Leibe, A. Leonardis, and B. Schiele. Robust Object Detection with Interleaved Categorization and Segmentation. In IJVC’05, revised in 2007.� B. Leibe and B. Schiele. Interleaved object categorization and
seg-mentation. In BMVC’03 � B. Leibe, A. Leonardis, B. Schiele, Combined Object Categorization and
Visual object recognition course Visual object recognition course 6777767777
� B. Leibe, A. Leonardis, B. Schiele, Combined Object Categorization and Segmentation with an Implicit Shape Model, ECCV’04 Workshop
� Space-Time Trajectory Estimation –B. Leibe, N. Cornelis, K. Cornelis, and L. Van Gool. Dynamic 3d scene analysis from a moving vehicle. In CVPR’07, 2007.(CVPR'07 Best Paper Award)
� Optimization –A. Leonardis, A. Gupta, and R. Bajcsy. Segmentation of range images as the search for geometric parametric models. IJCV, 14, 1995.
![Page 15: “Coupled Detection and Trajectory Estimation for ...daphna/course/student lectures/dmitri hanukaev.pdf · B. Leibe, N. Cornelis, K. Cornelis, and L. Van Gool. Dynamic 3d scene analysis](https://reader033.vdocuments.net/reader033/viewer/2022042000/5e6d6f9b2ec42c69fd334f30/html5/thumbnails/15.jpg)
�� IntroductionIntroduction�� Related WorksRelated Works�� Previous WorksPrevious Works
�� Pedestrian DetectionPedestrian Detection
Visual object recognition course Visual object recognition course 6777767777
�� SpaceSpace--Time Trajectory EstimationTime Trajectory Estimation
�� Coupling between Detection and Coupling between Detection and Trajectory EstimationTrajectory Estimation
�� ResultsResults
![Page 16: “Coupled Detection and Trajectory Estimation for ...daphna/course/student lectures/dmitri hanukaev.pdf · B. Leibe, N. Cornelis, K. Cornelis, and L. Van Gool. Dynamic 3d scene analysis](https://reader033.vdocuments.net/reader033/viewer/2022042000/5e6d6f9b2ec42c69fd334f30/html5/thumbnails/16.jpg)
Detection modelDetection model
Constellation Model: Parts and StructureBastian Leibe, Aleˇs Leonardis, and Bernt SchieleRobust Object Detection with Interleaved Categorization
and Segmentation, in IJCV 2005.
Visual object recognition course 67777Visual object recognition course 67777
Image adopted from the course slides
![Page 17: “Coupled Detection and Trajectory Estimation for ...daphna/course/student lectures/dmitri hanukaev.pdf · B. Leibe, N. Cornelis, K. Cornelis, and L. Van Gool. Dynamic 3d scene analysis](https://reader033.vdocuments.net/reader033/viewer/2022042000/5e6d6f9b2ec42c69fd334f30/html5/thumbnails/17.jpg)
Detection model Detection model -- TrainingTraining
Implicit Shape Model - ISM
Visual object recognition course Visual object recognition course 6777767777
Image adopted from B.Leibe
![Page 18: “Coupled Detection and Trajectory Estimation for ...daphna/course/student lectures/dmitri hanukaev.pdf · B. Leibe, N. Cornelis, K. Cornelis, and L. Van Gool. Dynamic 3d scene analysis](https://reader033.vdocuments.net/reader033/viewer/2022042000/5e6d6f9b2ec42c69fd334f30/html5/thumbnails/18.jpg)
Pedestrian detection Pedestrian detection ––Hypothesis’ Building ExampleHypothesis’ Building Example
Visual object recognition course Visual object recognition course 6777767777
Image adopted from B.Leibe
![Page 19: “Coupled Detection and Trajectory Estimation for ...daphna/course/student lectures/dmitri hanukaev.pdf · B. Leibe, N. Cornelis, K. Cornelis, and L. Van Gool. Dynamic 3d scene analysis](https://reader033.vdocuments.net/reader033/viewer/2022042000/5e6d6f9b2ec42c69fd334f30/html5/thumbnails/19.jpg)
MDL Hypothesis selection MDL Hypothesis selection ––in Generalin General
� ProblemWe have an over-complete set of hypothetical models. Who is the best?
� IntuitionTo prefer simple explanations to more complicated
Visual object recognition course Visual object recognition course 6777767777
To prefer simple explanations to more complicated ones.
� SolutionMinimum Description Length (MDL) (Rissanen 1984):the best encoding (model representation) is the one that minimizes the total description length for image, model, and error.
![Page 20: “Coupled Detection and Trajectory Estimation for ...daphna/course/student lectures/dmitri hanukaev.pdf · B. Leibe, N. Cornelis, K. Cornelis, and L. Van Gool. Dynamic 3d scene analysis](https://reader033.vdocuments.net/reader033/viewer/2022042000/5e6d6f9b2ec42c69fd334f30/html5/thumbnails/20.jpg)
MDL Hypothesis selection MDL Hypothesis selection --SolutionSolution
� Sdata – number N of data points, which are explained by H .
� Smodel denotes the cost of coding the model itself.
1 mod 2~h data el errorS S k S k S− −
Visual object recognition course 67777Visual object recognition course 67777
� Serror describes the cost for the error committed by the representation.
� κ1, κ2 are constants to weight the different factors
![Page 21: “Coupled Detection and Trajectory Estimation for ...daphna/course/student lectures/dmitri hanukaev.pdf · B. Leibe, N. Cornelis, K. Cornelis, and L. Van Gool. Dynamic 3d scene analysis](https://reader033.vdocuments.net/reader033/viewer/2022042000/5e6d6f9b2ec42c69fd334f30/html5/thumbnails/21.jpg)
MDL Hypothesis selectionMDL Hypothesis selection ––Solution of Quadratic Boolean ProblemSolution of Quadratic Boolean Problem� Optimal set of models :
� n = [n1, n2,..., nN] is a vector of indicator variables, such that ni =
11 1
1
. .
. . .max ,
. . .
. .
N
T
N NN
s s
n Sn S
s s
=
Visual object recognition course Visual object recognition course 6777767777
� n = [n1, n2,..., nN] is a vector of indicator variables, such that ni = 1 if hypothesis hi is accepted, and ni = 0 otherwise.
� S is an interaction matrix.� diagonal elements sii are the merit terms, individual hypotheses.� the off-diagonal elements (sij + sji) express the interaction costs
between two hypotheses hi and hj.
![Page 22: “Coupled Detection and Trajectory Estimation for ...daphna/course/student lectures/dmitri hanukaev.pdf · B. Leibe, N. Cornelis, K. Cornelis, and L. Van Gool. Dynamic 3d scene analysis](https://reader033.vdocuments.net/reader033/viewer/2022042000/5e6d6f9b2ec42c69fd334f30/html5/thumbnails/22.jpg)
Implementation of MDL for Implementation of MDL for DetectionDetection
Visual object recognition course Visual object recognition course 6777767777
Constraint:
Each pixel may at most belong to a single detection.
Image adopted from B.Leibe
![Page 23: “Coupled Detection and Trajectory Estimation for ...daphna/course/student lectures/dmitri hanukaev.pdf · B. Leibe, N. Cornelis, K. Cornelis, and L. Van Gool. Dynamic 3d scene analysis](https://reader033.vdocuments.net/reader033/viewer/2022042000/5e6d6f9b2ec42c69fd334f30/html5/thumbnails/23.jpg)
�� IntroductionIntroduction�� Related WorksRelated Works�� Previous WorksPrevious Works
�� Pedestrian DetectionPedestrian Detection
Visual object recognition course Visual object recognition course 6777767777
�� SpaceSpace--Time Trajectory EstimationTime Trajectory Estimation
�� Coupling between Detection and Coupling between Detection and Trajectory EstimationTrajectory Estimation
�� ResultsResults
![Page 24: “Coupled Detection and Trajectory Estimation for ...daphna/course/student lectures/dmitri hanukaev.pdf · B. Leibe, N. Cornelis, K. Cornelis, and L. Van Gool. Dynamic 3d scene analysis](https://reader033.vdocuments.net/reader033/viewer/2022042000/5e6d6f9b2ec42c69fd334f30/html5/thumbnails/24.jpg)
SpaceSpace--Time Trajectory EstimationTime Trajectory Estimation
t
� Trajectory growing� Collect detections in event cone� Evaluate under trajectory
Visual object recognition course Visual object recognition course 6777767777
x
t
z
itiH ,
Image adopted from B.Leibe
![Page 25: “Coupled Detection and Trajectory Estimation for ...daphna/course/student lectures/dmitri hanukaev.pdf · B. Leibe, N. Cornelis, K. Cornelis, and L. Van Gool. Dynamic 3d scene analysis](https://reader033.vdocuments.net/reader033/viewer/2022042000/5e6d6f9b2ec42c69fd334f30/html5/thumbnails/25.jpg)
SpaceSpace--Time Trajectory EstimationTime Trajectory Estimation
t
� Trajectory growing� Collect detections in event cone� Evaluate under trajectory� Adapt trajectory
Visual object recognition course Visual object recognition course 6777767777
x
t
z
itiH ,
![Page 26: “Coupled Detection and Trajectory Estimation for ...daphna/course/student lectures/dmitri hanukaev.pdf · B. Leibe, N. Cornelis, K. Cornelis, and L. Van Gool. Dynamic 3d scene analysis](https://reader033.vdocuments.net/reader033/viewer/2022042000/5e6d6f9b2ec42c69fd334f30/html5/thumbnails/26.jpg)
SpaceSpace--Time Trajectory EstimationTime Trajectory Estimation
t
� Trajectory growing� Collect detections in event cone� Evaluate under trajectory� Adapt trajectory� Iterate
Visual object recognition course 67777Visual object recognition course 67777
x
t
z
itiH ,
![Page 27: “Coupled Detection and Trajectory Estimation for ...daphna/course/student lectures/dmitri hanukaev.pdf · B. Leibe, N. Cornelis, K. Cornelis, and L. Van Gool. Dynamic 3d scene analysis](https://reader033.vdocuments.net/reader033/viewer/2022042000/5e6d6f9b2ec42c69fd334f30/html5/thumbnails/27.jpg)
� Trajectory growing� Collect detections in event cone� Evaluate under trajectory� Adapt trajectory� Iterate� Setting set as hypothesis
SpaceSpace--Time Trajectory EstimationTime Trajectory Estimation
t
Visual object recognition course Visual object recognition course 6777767777
� Setting set as hypothesis
x
t
z
itiH ,
![Page 28: “Coupled Detection and Trajectory Estimation for ...daphna/course/student lectures/dmitri hanukaev.pdf · B. Leibe, N. Cornelis, K. Cornelis, and L. Van Gool. Dynamic 3d scene analysis](https://reader033.vdocuments.net/reader033/viewer/2022042000/5e6d6f9b2ec42c69fd334f30/html5/thumbnails/28.jpg)
Implementation of MDL for Implementation of MDL for Trajectory EstimationTrajectory Estimation
� Trajectory selection� Start search from each detection.� Collect all resulting trajectories.� Perform hypothesis selection by
Visual object recognition course 67777Visual object recognition course 67777
solution of Quadratic Boolean Problem.
11 1
1
. .
. . .max ,
. . .
. .
M
T
M MM
q q
m Qm Q
q q
=
![Page 29: “Coupled Detection and Trajectory Estimation for ...daphna/course/student lectures/dmitri hanukaev.pdf · B. Leibe, N. Cornelis, K. Cornelis, and L. Van Gool. Dynamic 3d scene analysis](https://reader033.vdocuments.net/reader033/viewer/2022042000/5e6d6f9b2ec42c69fd334f30/html5/thumbnails/29.jpg)
Implementation of MDL for Implementation of MDL for Trajectory EstimationTrajectory Estimation
Visual object recognition course 67777Visual object recognition course 67777
Constraint:� Each detection can at most belong to a single
trajectory.� No two trajectories may intersect at any point in time.
Image adopted from B.Leibe
![Page 30: “Coupled Detection and Trajectory Estimation for ...daphna/course/student lectures/dmitri hanukaev.pdf · B. Leibe, N. Cornelis, K. Cornelis, and L. Van Gool. Dynamic 3d scene analysis](https://reader033.vdocuments.net/reader033/viewer/2022042000/5e6d6f9b2ec42c69fd334f30/html5/thumbnails/30.jpg)
�� IntroductionIntroduction�� Related WorksRelated Works�� Previous WorksPrevious Works
�� Pedestrian DetectionPedestrian Detection
Visual object recognition course Visual object recognition course 6777767777
�� SpaceSpace--Time Trajectory EstimationTime Trajectory Estimation
�� Coupling between Detection and Coupling between Detection and Trajectory EstimationTrajectory Estimation
�� ResultsResults
![Page 31: “Coupled Detection and Trajectory Estimation for ...daphna/course/student lectures/dmitri hanukaev.pdf · B. Leibe, N. Cornelis, K. Cornelis, and L. Van Gool. Dynamic 3d scene analysis](https://reader033.vdocuments.net/reader033/viewer/2022042000/5e6d6f9b2ec42c69fd334f30/html5/thumbnails/31.jpg)
Coupling between Detection and Coupling between Detection and Trajectory EstimationTrajectory Estimation
� Basic idea:� Couple the two optimization problems into a single one.� Move support for current detections into coupling terms.
� Coupling terms:� Express support for certain trajectories from new detections.� Express spatial prior for detection locations from trajectories.
Visual object recognition course 67777Visual object recognition course 67777
� Express spatial prior for detection locations from trajectories.
Image adopted from B.Leibe
![Page 32: “Coupled Detection and Trajectory Estimation for ...daphna/course/student lectures/dmitri hanukaev.pdf · B. Leibe, N. Cornelis, K. Cornelis, and L. Van Gool. Dynamic 3d scene analysis](https://reader033.vdocuments.net/reader033/viewer/2022042000/5e6d6f9b2ec42c69fd334f30/html5/thumbnails/32.jpg)
Coupling between Detection and Coupling between Detection and Trajectory EstimationTrajectory Estimation
� The Problem: Asymmetric relationship
� Trajectories rely on continuing detections for support.� But detections can exist without supporting
trajectories (e.g. when a new object enters the
Visual object recognition course 67777Visual object recognition course 67777
scene).
![Page 33: “Coupled Detection and Trajectory Estimation for ...daphna/course/student lectures/dmitri hanukaev.pdf · B. Leibe, N. Cornelis, K. Cornelis, and L. Van Gool. Dynamic 3d scene analysis](https://reader033.vdocuments.net/reader033/viewer/2022042000/5e6d6f9b2ec42c69fd334f30/html5/thumbnails/33.jpg)
Coupling between Detection and Coupling between Detection and Trajectory EstimationTrajectory Estimation
� Solution: Inserting of virtual trajectories v with interaction matrix R.� Enable detections to survive without contributing to an existing
trajectory
Visual object recognition course Visual object recognition course 6777767777
� W - Interaction between detections and virtual trajectories� V - Interaction between detections and real trajectories� U - Mutual exclusion between the two groups
Image adopted from B.Leibe
![Page 34: “Coupled Detection and Trajectory Estimation for ...daphna/course/student lectures/dmitri hanukaev.pdf · B. Leibe, N. Cornelis, K. Cornelis, and L. Van Gool. Dynamic 3d scene analysis](https://reader033.vdocuments.net/reader033/viewer/2022042000/5e6d6f9b2ec42c69fd334f30/html5/thumbnails/34.jpg)
�� IntroductionIntroduction�� Related WorksRelated Works�� Previous WorksPrevious Works
�� Pedestrian DetectionPedestrian Detection
Visual object recognition course 67777Visual object recognition course 67777
�� SpaceSpace--Time Trajectory EstimationTime Trajectory Estimation
�� Coupling between Detection and Coupling between Detection and Trajectory EstimationTrajectory Estimation
�� ResultsResults
![Page 35: “Coupled Detection and Trajectory Estimation for ...daphna/course/student lectures/dmitri hanukaev.pdf · B. Leibe, N. Cornelis, K. Cornelis, and L. Van Gool. Dynamic 3d scene analysis](https://reader033.vdocuments.net/reader033/viewer/2022042000/5e6d6f9b2ec42c69fd334f30/html5/thumbnails/35.jpg)
Results and DiscussionResults and Discussion
� Videos:
� Advantages
Visual object recognition course 67777Visual object recognition course 67777
� Limitations
� Conclusions
![Page 36: “Coupled Detection and Trajectory Estimation for ...daphna/course/student lectures/dmitri hanukaev.pdf · B. Leibe, N. Cornelis, K. Cornelis, and L. Van Gool. Dynamic 3d scene analysis](https://reader033.vdocuments.net/reader033/viewer/2022042000/5e6d6f9b2ec42c69fd334f30/html5/thumbnails/36.jpg)
ReferencesReferences
� B. Leibe, K. Schindler, L. Van Gool Coupled Detection and Trajectory Estimation for Multi-Object Tracking. In ICCV'07
� B. Leibe, A. Leonardis, and B. Schiele. Robust Object Detection with Interleaved Categorization and Segmentation. In IJVC’05, revised in 2007.
� B. Leibe, E. Seemann, and B. Schiele. Pedestrian detection in crowded scenes. In CVPR’05.
� B. Leibe and B. Schiele. Interleaved object categorization and seg-mentation. In BMVC’03.
Visual object recognition course Visual object recognition course 6777767777
seg-mentation. In BMVC’03.� B. Leibe, A. Leonardis, B. Schiele, Combined Object Categorization and
Segmentation with an Implicit Shape Model, ECCV’04 Workshop.� B. Leibe, N. Cornelis, K. Cornelis, and L. Van Gool. Dynamic 3d scene
analysis from a moving vehicle. In CVPR’07. (CVPR'07 Best Paper Award)� A. Leonardis, A. Gupta, and R. Bajcsy. Segmentation of range images as
the search for geometric parametric models. IJCV, 14, 1995.� K. Okuma, A. Taleghani, N. de Freitas, J. Little, and D. Lowe.
A boosted particle filter: Multi-target detection and tracking. In ECCV’04
![Page 37: “Coupled Detection and Trajectory Estimation for ...daphna/course/student lectures/dmitri hanukaev.pdf · B. Leibe, N. Cornelis, K. Cornelis, and L. Van Gool. Dynamic 3d scene analysis](https://reader033.vdocuments.net/reader033/viewer/2022042000/5e6d6f9b2ec42c69fd334f30/html5/thumbnails/37.jpg)
Questions…Questions…
Visual object recognition course Visual object recognition course 6777767777