Effective Local and Global Search for Fast Long-Term Tracking.

Area of Science:

• Computer Vision
• Artificial Intelligence

Background:

• Long-term object tracking is a challenging computer vision task, often limited by existing algorithms.
• Current methods struggle with maintaining target identity over extended periods and large image regions.

Purpose of the Study:

• To develop a novel, robust, and real-time framework for long-term object tracking.
• To improve tracking performance by integrating effective local search and global re-detection strategies.

Main Methods:

• A local search module with a bounding box regressor and a target verifier using a long short-term updated scheme.
• A novel re-detection module to estimate target position and size, adaptable to various base trackers.
• A strategy to dynamically switch between local and global search based on target presence verification.

Main Results:

• The proposed framework achieves satisfactory performance on popular datasets (VOT, OxUvA, TLP, LaSOT).
• The re-detection module demonstrates flexibility and effectiveness when integrated into other tracking algorithms.
• The tracker operates at real-time speeds while maintaining high accuracy.

Conclusions:

• The novel framework offers a significant advancement in robust and real-time long-term object tracking.
• The proposed local search and re-detection modules effectively address limitations of previous tracking methods.
• The framework provides a valuable tool for various computer vision applications requiring persistent object tracking.