Robust Visual Tracking Using Sparse Discriminative Graph Embedding

Jidong Zhao, Jingjing Li, Ke LuIEICE Trans. Information and Systems 2015 · OpenReview

Introduction

Visual tracking is a fundamental task in computer vision, yet it faces persistent challenges such as drastic changes in illumination, 3D pose variations, and scale fluctuations. Two primary issues plague traditional subspace learning methods like PCA or LDA: the inability to capture multi-modal appearance distributions and the "drift" problem, where small errors accumulate until the tracker loses the target to background clutter. This paper proposes Discriminative Graph Embedding Tracking (DGET), a framework that integrates graph-based subspace learning with sparsity constraints to address these limitations.

Proposed Methodology

The DGET algorithm is built upon three core components: a discriminative graph structure, sparse embedding via $L_{2,1}$-norm regularization, and an incremental update scheme.

Discriminative Graph Structure

To handle multi-modality and background clutter, the authors construct two distinct graphs. The within-class graph connects various "views" or clusters of the target to preserve its local manifold structure. Conversely, the between-class graph explicitly models the relationship between the target and the surrounding background, maximizing their separation in the feature space.

Conceptual overview of the discriminative graph structure showing foreground (circles) and background (triangles) separation across different stages of modeling.

Sparse Embedding and Feature Selection

The model employs $L_{2,1}$-norm regularization to achieve a "row-sparse" projection matrix. This approach performs simultaneous dimensionality reduction and feature selection, discarding irrelevant features that do not contribute to distinguishing the target from the background. This sparsity ensures both computational efficiency and robustness against noise.

Incremental Updating and Bayesian Inference

The tracker adapts to appearance changes through an incremental update scheme. It prunes old, irrelevant samples based on a "local manifold criterion"—specifically, removing samples with the fewest degrees in the graph—and incorporates new samples. This process is integrated into a Bayesian framework using a compound motion model to handle both smooth and abrupt movements.

Key Insights

A central insight of this work is the emphasis on local manifold structure. Unlike traditional LDA, which relies on global statistics, DGET respects the local neighborhood of data points. This is crucial for tracking objects that undergo non-linear appearance changes.

Illustration of the local manifold structure and the connections between different appearance clusters within the target class.

Furthermore, the use of the $L_{2,1}$-norm treats subspace learning as an automated feature selection process, identifying the most discriminative "parts" of the target image to maintain tracking stability.

Experimental Evaluation

The DGET algorithm was evaluated against state-of-the-art trackers like Incremental Visual Tracking (IVT) and Superpixel Tracking (ST) across several benchmarks:

  • Dudek & David: DGET demonstrated superior performance in handling 3D pose changes, scaling, and significant illumination shifts.
  • Faceocc2: The tracker remained robust during severe partial occlusions, where competing methods like IVT often failed when the target was obscured.
  • Woman & Singer: These sequences confirmed DGET's ability to resist background clutter and handle complex environments.

Quantitatively, DGET maintained a significantly lower center error in the Dudek sequence over 600 frames compared to IVT and Locally Connected Graph (LCG) models.

Conclusion and Future Work

The DGET framework provides a robust solution for visual tracking by combining discriminative graph embedding with sparse feature selection. While highly effective for targets with clear manifold structures (like faces), future work aims to extend this approach to highly deformable objects and reduce the reliance on initial labeled samples, moving toward a "zero-shot" updating strategy.

Summary updated

BibTeX
@article{zhao2015robust,
  author = {Jidong Zhao and Jingjing Li and Ke Lu},
  title = {{Robust Visual Tracking Using Sparse Discriminative Graph Embedding}},
  journal = {IEICE Trans. Information and Systems},
  year = {2015},
  url = {https://openreview.net/forum?id=AHa6sxMHNo},
}
Figure