Loading...

Object Tracking Via Sparse Representation Model

Zarezade, Ali | 2013

1241 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: Farsi
  3. Document No: 44772 (19)
  4. University: Sharif University of Technology
  5. Department: Computer Engineering
  6. Advisor(s): Rabiee, Hamid Reza
  7. Abstract:
  8. Visual tracking is a classic problem, but is continuously an active area of research, in computer vision. Given a bounding box defining the object of interest (target) in the first frame of a video sequence, the goal of a general tracker is to determine the ob-ject’s bounding box in subsequent frames. Utilizing sparse representation, we propose a robust tracking algorithm to handle challenges such as illumination variation, pose change, and occlusion. Object appearance is modeled using a dictionary composed of target patch images contained in previous frames. In each frame, the target is found from a set of candidates via a likelihood measure that is proportional to the sum of the reconstruction error of each candidate’s patch. Since the object appearance changes slowly in a video sequence, we expect the target in the current frame and the best can-didates of previous frames to belong to the same subspace. We impose this intuition by using joint sparsity inducing norms, to enforce the target and previous best candidates to have the same sparsity pattern. Moreover, the dictionary is updated in a patchwise manner. Occlusion state of patches is approximated by using patch reconstruction er-ror, and a prior occlusion probability found from a Markov chain. In each frame, the non-occluded patches of the best candidate are replaced by corresponding patches of a previously found target image in the dictionary. Extensive experimental results on sev-eral challenging sequences shows that the proposed method outperforms state-of-the-art
    trackers
  9. Keywords:
  10. Markov Chain ; Visual Tracking ; Particle Filter ; Sparse Representation ; Dictionary Learning