Unified Graph-Based Multicue Feature Fusion for Robust Visual Tracking

Area of Science:

• Computer Vision
• Artificial Intelligence

Background:

• Visual tracking faces challenges from unconstrained appearance variations and dynamic environments.
• Effective extraction of complementary information and adaptation to target appearance changes are critical.

Purpose of the Study:

• Propose a robust object tracking framework using unified graph fusion (UGF) of multicue.
• Address the key problems of feature deficiency and appearance variation in visual tracking.

Main Methods:

• Developed a unified graph fusion (UGF) framework integrating sparse and dense features via cross-diffusion.
• Implemented an iterative process to build unified features invariant to deformations, fast motion, and occlusion.
• Introduced a kernel-based adaptation strategy with outlier detection and a transductive reliability metric for appearance model updates.

Main Results:

• The proposed UGF tracker demonstrates robustness against object deformations, fast motion, and occlusion.
• Unified features enhance foreground-background discrimination, improving resilience to background clutter.
• The tracker favorably performs against 18 state-of-the-art trackers on benchmark datasets (OTB-50, OTB-100, VOT2017/18, UAV123).

Conclusions:

• The UGF framework effectively integrates multicue information for robust visual tracking.
• The novel adaptation strategy enhances the tracker's ability to handle scale, illumination, and rotation variations.
• The proposed tracker offers superior performance in challenging object tracking scenarios.