MAT: Multianchor Visual Tracking With Selective Search Region.

Area of Science:

• Computer Vision
• Machine Learning
• Robotics

Background:

• Modern visual trackers rely on motion assumptions, predicting current object locations based on previous predictions.
• Complex natural scenes and fast motion can lead to inaccurate predictions, causing tracking drift.

Purpose of the Study:

• To propose a real-time multianchor visual tracking (MAT) mechanism to enhance tracking anchor reliability.
• To overcome the limitations of relying solely on previous predictions for tracking anchors.

Main Methods:

• MAT selects the optimal tracking anchor from an ensemble, including objectness-based proposals and the previous prediction.
• An entropy-minimization method is used to select the best anchor from the ensemble.
• This approach utilizes selective search regions to improve tracking success rates.

Main Results:

• MAT enhances the reliability of tracking anchors, mitigating issues caused by inaccurate predictions.
• The method increases the probability of successful tracking with manageable computational costs.
• Experiments show MAT effectively improves the performance of nine base trackers across four challenging datasets.

Conclusions:

• Multianchor tracking (MAT) offers a robust solution for visual object tracking in challenging environments.
• The proposed anchor selection mechanism significantly improves tracking accuracy and stability.
• MAT provides a valuable enhancement for existing visual tracking algorithms.