Real-Time Visual Tracking through Fusion Features.

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Correlation filters are crucial for high-speed object tracking.
• Traditional trackers often rely on single feature types, limiting performance.
• Integrating multiple features presents challenges for existing correlation filter frameworks.

Purpose of the Study:

• To develop a novel approach for integrating multiple feature types in object tracking.
• To propose a multi-vector correlation filter (MVCF) capable of handling multi-vector descriptors.
• To enhance the performance and robustness of object tracking algorithms.

Main Methods:

• Introduction of a new DD-HOG (discriminative descriptors-histograms of oriented gradients) fusion feature.
• Development of a multi-vector correlation filter (MVCF) designed for direct convolution with multi-vector descriptors.
• Evaluation using the CVPR2013 tracking benchmark against state-of-the-art trackers.

Main Results:

• The proposed MVCF tracker demonstrates high effectiveness and speed.
• The DD-HOG descriptor integrated with MVCF shows superior performance compared to single-feature trackers.
• The MVCF tracker outperforms the structure-preserving object tracker (SPOT) in multi-object tracking, especially under heavy occlusion.

Conclusions:

• The MVCF provides an effective solution for utilizing multi-vector descriptors in correlation filter-based tracking.
• The DD-HOG fusion feature significantly improves tracking accuracy and robustness.
• The proposed method offers a promising advancement for high-speed and occluded object tracking applications.