Learning Feature Channel Weighting for Real-Time Visual Tracking.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Siamese convolutional neural networks (SCNNs) are effective for visual tracking but require extensive offline training.
• This offline training demands significant computational resources and time.
• Existing SCNNs often use features not optimized for tracking tasks.

Purpose of the Study:

• To improve visual tracking efficiency and reduce computational costs.
• To leverage pre-trained densely connected neural networks for robust feature extraction in visual tracking.
• To develop a novel method for selecting and fusing features suitable for real-time tracking.

Main Methods:

• Utilized a pre-trained densely connected neural network for initial feature extraction.
• Designed a regression network to determine the importance of each feature channel for target tracking.
• Proposed a weighting fusion strategy to select optimal features for visual tracking.

Main Results:

• The proposed channel weighting method demonstrated superior feature selection capabilities.
• Visualization of feature heatmaps confirmed the effectiveness of the channel weighting approach.
• Experimental results on four benchmarks showed the algorithm achieved state-of-the-art performance on key indicators.

Conclusions:

• The developed method enhances visual tracking efficiency and accuracy by intelligently utilizing pre-trained deep features.
• The channel weighting strategy effectively adapts general classification features for specific tracking tasks.
• This approach offers a computationally efficient and high-performing solution for visual tracking.