Orthogonal Single-Target Tracking

Youjin Kim, Junseok KwonIEEE Access 2022 · DOI

Abstract

Visual tracking remains a challenging task due to occlusions, rapid motion, and appearance variations. This paper presents a novel family of trackers based on the Wasserstein distance, utilizing Orthogonal Monte Carlo sampling to achieve high accuracy and real-time performance. By modeling target appearances as distributions and employing efficient projection techniques, the proposed Orthogonal Coupled Wasserstein-based Tracker (OCWT) achieves state-of-the-art results on major benchmarks while maintaining a processing speed of 79 FPS.

Introduction

The core of visual tracking is the ability to accurately match a target's appearance across consecutive video frames. Traditional methods often rely on $L_1/L_2$ norms or Kullback-Leibler (KL) divergence. However, these metrics frequently fail to capture the complex geometric relationships within appearance distributions. KL divergence, for instance, is asymmetric and can become discontinuous, leading to tracking failures during drastic appearance changes or occlusions.

The Wasserstein distance, derived from Optimal Transport theory, offers a more robust alternative by measuring the cost of transforming one distribution into another. While theoretically superior, its high computational complexity has historically limited its use in real-time tracking. This research addresses this limitation through efficient distributional modeling and sampling strategies.

Methodology: From Slicing to Orthogonal Coupling

The authors propose a hierarchical approach to integrating Wasserstein distance into tracking, culminating in the Orthogonal Coupled Wasserstein-based Tracker (OCWT).

Overview of the proposed tracking framework using Wasserstein distance metrics.

The tracking process begins by representing the target appearance as a distribution of statistical moments, including mean, variance, skewness, and kurtosis. This distributional modeling allows the tracker to better handle uncertainty compared to traditional point-based feature vectors. To compute the distance between the reference and estimated distributions efficiently, three methods are explored:

  1. Sliced Wasserstein-based Tracker (SWT): This method projects high-dimensional distributions onto multiple 1D lines (slices). In 1D, the Wasserstein distance is computationally trivial, allowing for rapid estimation.
  2. Projected Wasserstein-based Tracker (PWT): To mitigate the bias introduced by simple slicing, PWT uses bijective mapping functions to preserve the spatial ordering of points during projection, improving accuracy.
  3. Orthogonal Coupled Wasserstein-based Tracker (OCWT): The primary contribution of this work. Instead of random projections, OCWT utilizes Orthogonal Monte Carlo (OMC) sampling. By ensuring that projection vectors are mutually perpendicular, the system samples the appearance space more comprehensively with fewer samples, reducing variance and enhancing tracking stability.

Experimental Results

The proposed trackers were evaluated on three benchmark datasets: OTB, LaSOT, and VOT2017.

Accuracy and Success Rate

On the OTB dataset, OCWT achieved a success rate (AUC) of 0.710, outperforming established deep trackers such as SiamRPN++ (0.691) and ECO (0.709). Its performance on the long-term LaSOT dataset (AUC of 0.523) further demonstrates its stability over extended sequences, surpassing GlobalTrack and ATOM.

Real-Time Performance

A critical advantage of the OCWT is its efficiency. Despite the mathematical complexity of Optimal Transport, the sliced and orthogonal projection approach allows the tracker to operate at 79 FPS. This speed is achieved using a standard VGG backbone, making it significantly faster than many complex deep-learning architectures.

Conclusion

This paper demonstrates that the Wasserstein distance, when combined with Orthogonal Monte Carlo sampling, provides a powerful and efficient framework for visual tracking. By enforcing orthogonality in projections, OCWT maximizes information gain from limited samples, leading to a robust tracker that balances accuracy and speed. Future work may involve integrating this distributional matching logic into transformer-based architectures to further enhance tracking capabilities.

Summary updated

BibTeX
@article{kim2022orthogonal,
  author = {Youjin Kim and Junseok Kwon},
  title = {{Orthogonal Single-Target Tracking}},
  journal = {IEEE Access},
  year = {2022},
  doi = {10.1109/access.2022.3162200},
}
Figure