A unified decision-driven framework: Long-term tracking via visual-language models with motion estimation

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Current vision-language multimodal long-term tracking methods require extensive data and complex models, leading to high computational costs.
• This limits their practical application, especially in environments with limited computational resources.

Purpose of the Study:

• To propose a training-free, fast-response tracking framework (VL-METrack) that overcomes the limitations of existing methods.
• To transform long-term tracking from a data-driven problem to an adaptive decision-making process.

Main Methods:

• VL-METrack dynamically fuses motion estimation and localization networks.
• It utilizes a collaborative mechanism of optical flow and correlation filters for ambiguous tracking scenarios.
• A dynamic template update strategy and global re-detection mechanism are employed for maintaining template timeliness and optimal positioning.

Main Results:

• VL-METrack achieves performance comparable to state-of-the-art methods on OTB, LaSOT, and TNL2K datasets.
• The framework demonstrates significantly lower computational complexity and reduced resource consumption.
• It proves effective in resource-constrained scenarios.

Conclusions:

• VL-METrack offers an efficient and effective solution for vision-language multimodal long-term tracking.
• The training-free, adaptive approach makes it suitable for real-world applications with limited resources.
• This work advances the field by providing a computationally inexpensive yet high-performing tracking framework.