Lightweight Siamese Network with Global Correlation for Single-Object Tracking.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Siamese-based trackers have advanced object tracking but often lack efficiency.
• High precision in trackers can lead to increased complexity, limiting real-time applications.
• Resource-constrained environments require efficient and performant tracking solutions.

Purpose of the Study:

• To develop a novel lightweight Siamese network tracker (SiamGCN) for improved performance on devices with limited resources.
• To address the trade-off between tracking precision and computational complexity.
• To enhance feature integration for more robust object tracking.

Main Methods:

• Utilized MobileNet-V3 as a backbone for efficient feature extraction, with stride modifications.
• Developed a global correlation module based on Transformer architecture with multi-head cross-attention.
• Implemented global feature fusion to integrate template and search region information.

Main Results:

• SiamGCN demonstrated high tracking performance across VOT2018, VOT2019, LaSOT, and TrackingNet benchmarks.
• The proposed tracker significantly reduced parameter count and computational costs.
• Achieved substantial improvements in processing speed and resource utilization.

Conclusions:

• SiamGCN offers a compelling solution for efficient and accurate object tracking in resource-limited scenarios.
• The lightweight architecture and global feature fusion effectively balance performance and efficiency.
• This work contributes to the practical deployment of advanced object tracking technologies on edge devices.