Robust Visual Tracking Based on Adaptive Convolutional Features and Offline Siamese Tracker.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Robust visual tracking is a key challenge in computer vision.
• Existing methods like spatially regularized discriminative correlation filter (SRDCF) struggle with partial-target or background learning during occlusion or out-of-view events.
• Limited training data necessitates robust target appearance models.

Purpose of the Study:

• To develop a novel method for enhancing visual tracking ability.
• To reduce the computational complexity of existing tracking algorithms.
• To achieve robust and accurate long-term visual tracking.

Main Methods:

• Adaptive dimensionality reduction using pre-trained VGG-Net for feature extraction.
• Adaptive model update strategy based on peak-to-sidelobe ratio weighting.
• Integration of an online SRDCF-based tracker with an offline Siamese tracker for long-term tracking.

Main Results:

• The proposed tracker demonstrates satisfactory performance across various challenging tracking scenarios.
• The adaptive techniques contribute to improved tracking accuracy and robustness.
• Reduced computational complexity compared to existing methods.

Conclusions:

• The novel approach effectively addresses limitations in current visual tracking techniques.
• The combined tracker offers a robust solution for long-term tracking under difficult conditions.
• The method shows promise for real-world applications requiring reliable object tracking.