CVTrack: Combined Convolutional Neural Network and Vision Transformer Fusion Model for Visual Tracking

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Current single-object trackers predominantly use either Convolutional Neural Networks (CNNs) or Vision Transformers (ViTs) as their backbone.
• CNNs excel at local feature extraction but struggle with global representations.
• ViTs capture long-range dependencies via self-attention but may miss local details.

Purpose of the Study:

• To address the limitations of existing single-object tracking methods.
• To propose a novel tracking algorithm, CVTrack, that integrates both local and global feature extraction capabilities.
• To achieve state-of-the-art performance while maintaining real-time execution speed.

Main Methods:

• Introduced CVTrack, a target-tracking algorithm featuring a parallel dual-branch backbone network combining CNN and Transformer architectures.
• Employed bidirectional information interaction channels for effective fusion of local (CNN) and global (Transformer) features.
• Utilized deep cross-correlation and transformer-based methods for comprehensive template and search region feature fusion before prediction.

Main Results:

• CVTrack achieved state-of-the-art performance across five benchmark datasets.
• The proposed tracker demonstrated real-time execution capabilities.
• Ablation studies confirmed the effectiveness of individual modules within the parallel dual-branch backbone.

Conclusions:

• The parallel dual-branch backbone effectively leverages the strengths of both CNNs and Transformers for superior object tracking.
• CVTrack offers a robust solution for single-object tracking, balancing accuracy and speed.
• The fusion strategy significantly enhances the interaction between template and search region features.