LPCF: Robust Correlation Tracking via Locality Preserving Tracking Validation.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Online model updates in visual tracking are prone to incorporating corrupted training samples, leading to tracking failures.
• Model drift is a significant challenge in visual tracking, degrading performance over time.

Purpose of the Study:

• To develop a novel and generic decontamination approach to mitigate model drift in visual tracking.
• To enhance the robustness and accuracy of visual tracking algorithms by validating tracking results.

Main Methods:

• Integration of a locality preserving correlation filter (LPCF) with a new decontamination strategy.
• Modeling spatial neighborhood and topological structures of the bounding box center for tracking-result validation.
• Development of a closed-form solution for millisecond-level validation and a dimensionality reduction strategy for real-time performance.

Main Results:

• The decontamination approach significantly improves overall performance by 6.2% (OTB-2015), 12.6% (LASOT), and 3% (TrackingNet).
• The complete LPCF algorithm enhances baseline performance by 27.8% (OTB-2015), 34.8% (LASOT), and 15% (TrackingNet).
• The proposed tracker achieves top performance among existing decontamination trackers under real-time constraints.

Conclusions:

• The proposed decontamination approach effectively addresses model drift in visual tracking.
• The LPCF tracker offers a robust, accurate, and real-time solution for visual tracking challenges.
• The method demonstrates superior performance and efficiency compared to existing approaches.