Robust and Fast Collaborative Tracking with Two Stage

Robust and Fast Collaborative Tracking
with Two Stage Sparse Optimization
Authors: Baiyang Liu, Lin Yang, Junzhou Huang,
Peter Meer, Leiguang Gong and Casimir Kulikowski
Outline
 Problem of Tracking
 State of the art algorithms
 The proposed algorithm
 Experiment result
The problem
 Tracking: estimate the state of moving target in the observed
video sequences
 Challenges
 Illumination, pose of target changes
 Object occlusion, complex background clutters
 Landmark ambiguity
 Two categories of tracking
 Discriminative
 Generative
Outline
 Problem of Tracking
 State of the art algorithms
 The proposed algorithm
 Experiment result
Related work
 Multiple Instance Learning boosting method(MIL Boosting)
put all samples into bags and labeled them with bag labels.
 Incremental Visual Tracking(IVT)
the target is represented as a single online learned
appearance model
 L1 norm optimization
a linear combination of the learned template set composed of
both target templates and the trivial template.
Basic sparse representation
 Sparse representation
 Basis pursuit
 Disadvantages
 Computationally expensive
 Temporal and spatial features are not considered
 The background pixels do not lie on the linear template
subspace
Outline
 Problem of Tracking
 State of the art algorithms
 The proposed algorithm
 Experiment result
Problem Analysis
 Given
,Let
,
,
 Feature space can be decreased to K0 dimension
 Two stage greedy method
Stage I: Feature selection
 Loss function
Given
, L=
as labels,
 To minimize the loss function, solve the sparse problem
below
 Feature selection matrix
Stage II: Sparse reconstruction
 Problem after stage I
 Simplify the aim function above as
Bayesian tracking framework
 Let
represents the affine paramters
 Estimation of the state probability
prediction:
updating:
 Transition model:

~
likelihood
where
Review of the algorithm
Outline
 Problem of Tracking
 State of the art algorithms
 The proposed algorithm
 Experiment result
Visual results
Quantitative results