Learning Distance Constrained Transformation for Video Tracking in Car-Following

Area of Science:

• Computer Vision
• Machine Learning
• Robotics

Background:

• Discriminative correlation filters in video tracking often struggle with fusing diverse features, leading to performance degradation.
• Combining hand-crafted and deep learning features can cause resolution constraints, resulting in tracking errors like peak response slippage.

Purpose of the Study:

• To address inference conservatism in multi-type feature tracking.
• To improve the positional precision and robustness of video tracking algorithms, particularly for autonomous driving.

Main Methods:

• Developed a target-observation constraint framework to formalize discrimination conservatism across feature map channels.
• Introduced a learning constraint transformation methodology for clustering similar representations and separating dissimilar ones.
• Proposed an updating strategy to suppress low scores of symmetric dispersion ratio for enhanced robustness.

Main Results:

• The proposed framework significantly improves positional precision in detection responses through joint learning with correlation filters.
• The updating strategy enhances tracking robustness by effectively managing feature representations.
• Evaluations on five benchmark datasets (UAV20L, UAVDT, OTB-100, VOT-2019, LaSOT) show superior performance compared to existing methods.

Conclusions:

• The developed target-observation constraint framework offers a robust solution for multi-type feature video tracking.
• The approach enhances tracking accuracy and stability, making it suitable for demanding applications like autonomous driving.