Distortion-Aware Correlation Tracking.

Area of Science:

• Computer Vision
• Machine Learning
• Robotics

Background:

• Correlation filter (CF)-based tracking methods offer high-speed performance.
• Distortion in correlation outputs is a significant challenge for CF trackers, leading to tracking failures.

Purpose of the Study:

• To propose a distortion-aware correlation filter framework capable of detecting and recovering from tracking distortions.
• To enhance the robustness and reliability of CF-based trackers.

Main Methods:

• A normed correlation response feature is employed for distortion detection.
• A competition mechanism, formulated as a maximum multi clique problem on a specialized graph, handles distortions.
• A global-local context model is utilized to mitigate underlying distortions.

Main Results:

• The proposed tracker successfully identifies the optimal target trajectory during distortion periods.
• The framework demonstrates the ability to retrieve potentially lost targets.
• Experimental results on the Online Tracking Benchmark show superior performance compared to state-of-the-art methods.

Conclusions:

• The distortion-aware CF framework effectively addresses limitations in current CF trackers.
• The method significantly improves object tracking accuracy and robustness, especially under challenging conditions.
• This work advances the state-of-the-art in high-speed visual tracking.