Parallel Correlation Filters for Real-Time Visual Tracking

Area of Science:

• Computer Vision
• Machine Learning
• Robotics

Background:

• Correlation filter-based methods excel in visual object tracking accuracy and speed.
• Existing methods struggle with significant appearance changes like deformation, illumination variation, and rotation.

Purpose of the Study:

• Propose a novel parallel correlation filters (PCF) framework for robust real-time visual object tracking.
• Enhance tracking performance by addressing limitations of current correlation filter methods.

Main Methods:

• Construct two parallel correlation filters: one for appearance changes, one for translation.
• Employ weighted merging of response maps for accurate target center localization.
• Utilize a new correlation output distribution during training to improve filter accuracy and prevent model drift.

Main Results:

• The proposed PCF tracker demonstrates superior performance compared to state-of-the-art trackers.
• Achieves high real-time tracking performance on OTB-2013 and OTB-2015 benchmarks.
• Successfully handles target deformation, illumination variation, and rotation.

Conclusions:

• The PCF framework offers a robust and efficient solution for real-time visual object tracking.
• The novel approach significantly improves tracking accuracy and stability.
• Validated through extensive experiments on standard object tracking datasets.