Visual Tracking: An Experimental Survey

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Object tracking is a challenging computer vision problem with numerous proposed algorithms.
• Existing tracker evaluations often use limited datasets, hindering comprehensive performance assessment.

Purpose of the Study:

• To systematically evaluate nineteen object trackers on a diverse dataset of 315 video fragments.
• To provide an objective analysis of tracker strengths and weaknesses under various challenging conditions.

Main Methods:

• Experimental evaluation of nineteen selected object trackers.
• Utilized 315 video fragments encompassing illumination changes, occlusion, clutter, and camera motion.
• Employed survival curves, Kaplan Meier statistics, and Grubs testing for objective assessment.

Main Results:

• Demonstrated the effectiveness of survival curves and statistical methods for objective tracker evaluation.
• Found the F-score to be as effective as object tracking accuracy (OTA) for performance measurement.
• Identified specific strengths and weaknesses of various trackers across diverse scenarios.

Conclusions:

• Systematic, large-scale evaluation provides crucial insights into object tracker performance.
• Objective metrics like survival curves and F-score are vital for reliable tracker assessment.
• This work offers a benchmark for future object tracking research and development.