1 / 32

Bag of Features Tracking

Bag of Features Tracking. Fan Yang, Huchuan Lu School of Information and Communication Engineering Dalian University of Technology, Dalian, CHINA Yen-Wei Chen College of Information Science and Engineering Ritsumeikan University, Kusatsu, JAPAN. Goal of our work. Related algorithms.

brock
Télécharger la présentation

Bag of Features Tracking

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Bag of Features Tracking Fan Yang, Huchuan Lu School of Information and Communication Engineering Dalian University of Technology, Dalian, CHINA Yen-Wei Chen College of Information Science and Engineering Ritsumeikan University, Kusatsu, JAPAN

  2. Goal of our work

  3. Related algorithms • Bag of Features • aims to find a set of representative features from a large dataset of images • an image is represented as the occurrence of such features • mainly used for object categorization

  4. Bag of Features (BoF) [Fei-Fei and Perona, CVPR05]

  5. Main contribution • we propose a visual tracking approach based on “bag of features” (BoF) algorithm which fuse patch-based approach and global template-based approach into a unified framework. • Besides, updating mechanism and result refinement scheme are included in BoF tracking. • Experiments demonstrate that our approach is robust in handling occlusion, scaling and rotation.

  6. Overview of our approach • Patch generation • Codebook construction • Tracking • Updating • Results refinement

  7. Overview of our approach • Patch generation • Codebook construction • Tracking • Updating • Results refinement

  8. Patch generation Locate the object Randomly sample patches (constant size) Collection of patches

  9. Overview of our approach • Patch generation • Codebook construction • Tracking • Updating • Results refinement

  10. Codebook construction Accumulate training images Feature representation Clustering Formed codebook

  11. Codebook construction Note: We utilize RGB and LBP features to construct two codebooks separately

  12. Codebook construction We useM training frames, compute the frequency of occurrence of codewords within the tracked object region and obtain M histograms, called “bags”

  13. Overview of our approach • Patch generation • Codebook construction • Tracking • Updating • Results refinement

  14. Tracking For each new frame Randomly sample T candidates (particle sampling) Collection of candidates

  15. Tracking For each candidate (w.r.t. one codebook) sample patches extract features compute the minimum distance with Patch similarity computation

  16. Tracking For each candidate (w.r.t. one codebook) projection to the codebook compute the minimum distance with Bag similarity computation

  17. Tracking choose the best Using to weight and is the tracked object.

  18. Overview of our approach • Patch generation • Codebook construction • Tracking • Updating • Results refinement

  19. Updating • Collect p patches with highest value from the tracked object, using is predefined and constant • Repeat the above step in the following f frames to obtain patches • Perform K-means clustering again on and to get new codebook forget factor

  20. Updating new codebook

  21. Overview of our approach • Patch generation • Codebook construction • Tracking • Updating • Results refinement

  22. Results refinement When the similarity of BoF tracking is below a threshold , the result of BoF tracking is inaccurate Using the current and previous affine parameter to refine and is affine parameters of previous, current affine para and refined result

  23. Experimental results • Settings for BoF tracking • Features: 64-dim RGB and 59-dim LBP • Size of codebook: 20 • Number of sampled candidates: 300 • Number of patches in a candidate: 50 • Patch size: 12x12 pixels • Forget factor: = 0.9 • Training frames: 5 frames • Settings for comparison • IVT: 300 sampled candidates • FragTrack: 4 pixel or 5 pixel half size of search window IVT : Ross et al. Incremental learning for robust visual tracking. IJCV2007. FragTrack : Adam et al. Robust fragments-based tracking using the integral histogram. CVPR2006

  24. Occlusion Simple occlusion Continuous occlusion

  25. Background clutter

  26. Pose and scale change

  27. More results

  28. More results

  29. Influence of Parameters Patch size & weighting strategy Seq. / Param. 8×8 adaptive 12×12 adaptive 16×16 adaptive 20×20 adaptive 12×12 constant ShopAssistant2cor 65.03 3.36 4.02 14.84 5.28 OneStopEnter1front 5.65 4.01 5.37 9.24 4.28 occlusion1 11.23 8.23 9.31 12.10 31.78 boy 6.97 2.97 6.17 6.37 3.67 • average center location errors in pixels • “adaptive” means adaptive weight is used to combine RGB and LBP codebooks • “constant” means we use a constant weight (0.5 for both of the two codebooks)

  30. Summary • Bag of Features Tracking (BoF Tracking) • We incorporate BoF algorithm into tracking framework with effective modification • Robust in handling occlusion, pose changes and background clutter • Our method can be easily extended by adding spatial information

  31. Thank you! http://ice.dlut.edu.cn/lu/index.html

More Related