HFFTrack: Transformer tracking via hybrid frequency features.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Transformer-based trackers leverage global modeling but struggle with high-frequency features crucial for precise target localization.
• Existing methods often lack the capacity to effectively extract and utilize both high- and low-frequency target information.

Purpose of the Study:

• To develop an advanced object tracking algorithm that effectively fuses multi-frequency features for improved performance.
• To address the limitations of Transformer models in capturing high-frequency details essential for accurate tracking.

Main Methods:

• A novel feature extraction network combining Convolutional Neural Networks (CNN) and Transformers to learn multi-frequency target features in stages.
• A dual-branch encoder designed to capture both global context and local target features.
• A multi-frequency feature fusion network utilizing wavelet transform and convolution to integrate high- and low-frequency information.

Main Results:

• The proposed tracker demonstrates superior performance across six challenging benchmark datasets: LaSOT, TrackingNet, GOT-10k, TNL2K, UAV123, and OTB100.
• Experimental results validate the effectiveness of fusing hybrid frequency features for enhanced object tracking accuracy.

Conclusions:

• The hybrid frequency feature fusion approach significantly improves the performance of Transformer-based trackers.
• The designed network effectively balances high- and low-frequency information, leading to robust and accurate object tracking.