Parallel Correlation Filters for Real-Time Visual Tracking

Yijin Yang, Yihong Zhang, Demin Li, Zhijie Wang2019 · DOI

TL;DR

This paper introduces a novel "Parallel Correlation Filters" (PCF) framework for real-time visual object tracking. By employing two parallel correlation filters with different learning rates and a new "reasonable distribution" for correlation responses, PCF significantly improves tracking accuracy and robustness against appearance changes and model drift, outperforming state-of-the-art methods.

What Problem Does This Paper Address?

Visual tracking, the task of following an object in a video sequence, is a fundamental problem in computer vision with applications ranging from surveillance to autonomous driving. A major challenge in visual tracking is the unpredictable appearance changes of the target due to factors like deformation, illumination variations, and rotation. While discriminative correlation filter (DCF)-based methods have achieved impressive accuracy and speed, they often struggle with significant appearance changes, leading to model drift. Existing methods also tend to use a fixed learning rate for updating the target model, which is suboptimal. Furthermore, the assumption of a Gaussian distribution for correlation output can lead to inaccurate localization. This paper aims to address these limitations by proposing a more robust and accurate tracking framework.

Methods

The core of the proposed approach is the Parallel Correlation Filters (PCF) framework. It introduces two parallel correlation filters:

  • PCF1: Uses a larger learning rate to track rapid appearance changes in the target.
  • PCF2: Uses a smaller learning rate to track slower positional changes.

The response maps from these two filters are then weighted and fused to accurately estimate the target's center position.

To combat model drift and improve localization accuracy, the paper proposes a new "reasonable distribution" for the correlation response. This distribution organically merges the sharp peak of a Gaussian distribution with the steep hillside of a triangle distribution, creating a sharper peak and a larger slope. This new distribution is designed to prevent the model from drifting and enhance tracking robustness.

The framework also incorporates an adaptive update strategy where the two parallel filters are updated with different learning rates every six frames, allowing for better adaptation to varying target dynamics. For scale estimation, a standard DCF approach is utilized.

Datasets

The proposed PCF tracker is evaluated on two widely used benchmarks:

  • OTB-2013: A benchmark containing 50 sequences with 11 different attributes.
  • OTB-2015: A more challenging benchmark with 100 sequences and the same 11 attributes.

The paper does not introduce new datasets.

Key Insights and Ideas

  • Parallelism for Robustness: Using two correlation filters with different learning rates is a clever way to simultaneously handle both rapid appearance changes and slower positional shifts, a common challenge in tracking.
  • "Reasonable Distribution" for Sharpness: The novel correlation response distribution, merging Gaussian and triangle properties, is a key innovation. It addresses the limitations of smooth Gaussian peaks and shallow triangle slopes, leading to more precise localization and reduced drift.
  • Adaptive Learning Rates: The idea of using different learning rates for the parallel filters and updating them periodically (every six frames) allows the tracker to adapt more effectively to the target's behavior over time.

Key Results

The PCF tracker demonstrates superior performance on both OTB-2013 and OTB-2015 benchmarks.

  • Accuracy: PCF achieves state-of-the-art accuracy, outperforming many top trackers. For example, on OTB-2013, it achieves 89.2% precision and 67.5% success rate, a significant gain over the baseline.
  • Robustness: The tracker shows remarkable robustness against deformation, illumination variation, and rotation, as evidenced by its strong performance across various challenging attributes.
  • Real-time Performance: PCF achieves high tracking speeds, running at approximately 41-42 FPS on a CPU, making it suitable for real-time applications.

Limitations and Future Work

The paper acknowledges that the tracker can fail in cases of complete or heavy occlusions for extended periods. While deep learning-based trackers show more robustness in such scenarios, PCF can still drift.

As future work, the authors propose to:

  • Investigate new re-detection strategies to address issues with prolonged occlusions.
  • Analyze and potentially merge strategies from other works that handle occlusion and background clutter effectively.

Summary updated

BibTeX
@misc{yang2019parallel,
  author = {Yijin Yang and Yihong Zhang and Demin Li and Zhijie Wang},
  title = {{Parallel Correlation Filters for Real-Time Visual Tracking}},
  year = {2019},
  doi = {10.3390/s19102362},
}
Figure