A Visual Tracker Offering More Solutions.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Traditional visual trackers often prioritize robustness and accuracy over long-term sustainability and real-time processing.
• Addressing the limitations of existing methods, there is a need for advanced trackers that meet all these critical requirements.

Purpose of the Study:

• To propose a novel, state-of-the-art visual tracker that is robust, accurate, and capable of real-time processing for long-term tracking tasks.
• To enhance feature extraction, network stability, and target discrimination while optimizing computational efficiency.

Main Methods:

• Utilized EfficientNet-B0 as the backbone network via neural architecture search for improved feature extraction and reduced parameter count.
• Implemented maximal Distance Intersection-over-Union for target estimation, enhancing stability and convergence.
• Integrated channel and spatial dual attention mechanisms for improved target discrimination and conjugate gradient optimization for faster online learning.

Main Results:

• The proposed tracker demonstrates superior performance on benchmark datasets including OTB2015, VOT2016, VOT2018-LT, UAV-123, and LaSOT.
• Achieved real-time processing speeds exceeding 50 FPS, showcasing a significant speed advantage over pure global searching methods.
• The tracker effectively handles sudden target movements and brief disappearances through a proposed two-stage search and screening module.

Conclusions:

• The developed visual tracker offers a comprehensive solution by balancing robustness, accuracy, long-term sustainability, and real-time performance.
• The novel combination of EfficientNet-B0, advanced attention mechanisms, and optimized search strategies sets a new standard in visual tracking.
• This research holds significant practical value for applications requiring reliable and efficient real-time object tracking.