Robust online learning based on siamese network for ship tracking.

Area of Science:

• Computer Vision
• Machine Learning
• Robotics

Background:

• Inland river ship tracking is hindered by frequent occlusions, leading to inaccurate motion estimation, object tracking drift, and target loss.
• Existing tracking methods lack robustness in complex, dynamic environments with occlusions.

Purpose of the Study:

• To propose a robust online learning ship tracking algorithm to address challenges posed by occlusions in inland river scenes.
• To enhance the accuracy and reliability of ship motion state estimation in the presence of frequent occlusions.

Main Methods:

• A novel algorithm combining offline Siamese network and online classifier scores for discriminative learning.
• An occlusion determination mechanism based on a fused classification score to manage target template updates.
• Implementation of a global search mechanism for target re-localization during occlusion.
• Integration of an adaptive online update strategy, UpdateNet, to mitigate template degradation.

Main Results:

• The proposed algorithm demonstrates strong robustness in occlusion scenarios on inland river ship datasets.
• Achieved an accuracy rate of 56.8% and a success rate of 57.2% in challenging tracking conditions.
• Successfully avoids object tracking drift and target loss during occlusion events.

Conclusions:

• The developed Siamese network and region proposal network-based algorithm offers a significant improvement for robust ship tracking in occluded inland waterways.
• The adaptive online update strategy and occlusion handling mechanism contribute to enhanced tracking performance and reliability.