Part-Aware Framework for Robust Object Tracking.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Existing object tracking methods often rely on holistic target representations, making them sensitive to partial occlusion and appearance variations.
• Context regression trackers, while effective, can struggle with localized target changes.

Purpose of the Study:

• To propose a novel part-aware framework for robust object tracking.
• To address the limitations of holistic representations in existing context regression methods.
• To improve tracking performance in challenging scenarios like occlusion and drastic appearance changes.

Main Methods:

• A part-aware framework is proposed, integrating global and local target information for collaborative state awareness.
• A spatial-temporal measure is introduced to evaluate individual part regressor quality and balance global-local information.
• Part regressor outputs are aggregated using their measures as weights for refined target localization.
• Regressor divergence quantifies background noise interference, enabling adaptive filtering.
• Spatial-temporal information among part regressors aids in accurate target scale estimation.

Main Results:

• The proposed framework significantly enhances the performance of various context regression trackers.
• The method demonstrates competitive results against state-of-the-art trackers on multiple popular benchmarks.
• Evaluations were conducted on OTB, TC128, UAV, UAVDT, VOT, TrackingNet, GOT-10k, and LaSOT datasets.

Conclusions:

• The part-aware framework offers a robust solution for object tracking, particularly in scenarios with partial occlusion and appearance changes.
• Integrating local part information alongside global context improves tracking accuracy and resilience.
• The proposed approach provides a valuable enhancement for existing context regression-based object tracking systems.