Visual Tracking via Deep Feature Fusion and Correlation Filters

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Visual tracking is crucial for extracting semantic information from videos and images.
• Despite advancements, real-world tracking faces challenges like target pose changes, occlusion, and sudden movements, leading to target loss.

Purpose of the Study:

• To develop an effective visual tracking method that overcomes limitations of traditional artificial feature models.
• To improve tracker learning accuracy by leveraging deep hierarchical features from Convolutional Neural Networks (CNNs).

Main Methods:

• A hybrid tracker was developed, integrating deep feature extraction with correlation filter techniques.
• Multi-layer feature fusion from CNNs was employed to enhance the learning accuracy of the tracker.

Main Results:

• The proposed hybrid tracker demonstrated robust performance on benchmark datasets (OBT-100 and OBT-50).
• The method effectively addressed challenges such as significant pose variations and occlusions.

Conclusions:

• The hybrid deep feature and correlation filter tracker offers a powerful solution for challenging visual tracking tasks.
• The proposed algorithm shows significant effectiveness in improving tracking accuracy and robustness.