Learning Deep Lucas-Kanade Siamese Network for Visual Tracking.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Siamese trackers offer a balance of accuracy and efficiency in visual object tracking.
• Conventional Siamese trackers suffer from limited discriminative power due to insufficient template-candidate feature representation.
• Existing methods often use non-aligned features and neglect the exploration of target object's geometrical transformations.

Purpose of the Study:

• To propose a novel Siamese tracking framework that dynamically transforms template-candidate features for more discriminative similarity matching.
• To address the limitations of conventional Siamese trackers by incorporating geometrical transformation insights.

Main Methods:

• Reformulated template-candidate matching from the perspective of the Lucas-Kanade (LK) image alignment approach.
• Introduced a Lucas-Kanade network (LKNet) integrated into the Siamese architecture.
• LKNet learns aligned feature representations in a data-driven, trainable manner to enhance model adaptability.

Main Results:

• The proposed LKNet enhances feature representation for more discriminative template-candidate matching.
• The framework demonstrates improved adaptability in challenging tracking scenarios.
• Two Siamese trackers, LK-Siam and LK-SiamRPN, validated the effectiveness of the proposed approach.

Conclusions:

• The novel Siamese tracking framework with LKNet significantly improves visual tracking performance.
• The method proves more competitive than several state-of-the-art tracking approaches on prevalent datasets.
• The integration of LK image alignment principles enhances the robustness and accuracy of Siamese trackers.