Online Scale Adaptive Visual Tracking Based on Multilayer Convolutional Features.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Visual tracking is challenged by complex object appearances and occlusions.
• Convolutional Neural Networks (CNNs) offer powerful hierarchical feature extraction for tracking.
• Existing methods often struggle with scale variation and tracking drift.

Purpose of the Study:

• To propose an online scale adaptive tracking method leveraging multilayer CNN features.
• To enhance tracking accuracy and robustness against appearance changes and occlusion.
• To reduce computational complexity and mutual errors in translation and scale estimation.

Main Methods:

• Utilizing multilayer convolutional features jointly for robust feature representation.
• Implementing a scale pyramid for optimal and fast scale estimation.
• Employing separate correlation filters for translation and scale, with weighted fusion for localization.
• Introducing an adaptive and selective update mechanism for translation filters to combat drift.

Main Results:

• The proposed method demonstrates superior performance in visual tracking tasks.
• Effective handling of scale variations and target appearance changes was achieved.
• Reduced mutual errors between translation and scale estimations were observed.
• Improved robustness against severe occlusion was validated through experiments.

Conclusions:

• The proposed online scale adaptive tracking method offers excellent performance.
• Joint use of multilayer CNN features and separate correlation filters is effective.
• The adaptive update mechanism enhances tracking stability under challenging conditions.