Object Tracking Based on Vector Convolutional Network and Discriminant Correlation Filters.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Discriminant Correlation Filters (DCF) are popular for online object tracking due to speed and efficiency.
• Increasing performance in DCF methods often leads to higher parameter counts and reduced speed.
• Existing methods face challenges with drift, especially during fast motion.

Purpose of the Study:

• To propose a novel visual tracking algorithm, VDCFNet, that addresses the limitations of current DCF methods.
• To enhance tracking efficiency and robustness while maintaining low memory footprint.
• To improve generic feature learning and online tracking performance.

Main Methods:

• Integration of Discriminant Correlation Filter (DCF) with a Vector Convolutional Neural Network (VCNN).
• Replacement of traditional convolutional filters with two novel vector convolutional filters.
• Implementation of a coarse-to-fine search strategy for online tracking.
• Selective model updating for speed and robustness enhancement.

Main Results:

• VDCFNet demonstrates competitive performance on OTB benchmarks.
• The algorithm achieves real-time processing speeds.
• The model requires minimal memory (59 KB) after offline training.
• Improved robustness against drift, particularly in fast-motion scenarios.

Conclusions:

• VDCFNet offers an effective solution for real-time visual object tracking.
• The proposed VCNN integration and search strategy enhance tracking performance and efficiency.
• The algorithm presents a favorable trade-off between performance, speed, and memory usage.