Meta-Knowledge Learning and Domain Adaptation for Unseen Background Subtraction.

Area of Science:

• Computer Vision
• Machine Learning
• Video Processing

Background:

• Background subtraction is crucial for video analysis but deep learning models struggle with unseen scenarios.
• Existing methods lack robustness and generalization capabilities for diverse real-world applications.

Purpose of the Study:

• To develop a robust and generalizable background subtraction framework for unseen scenarios.
• To address the limitations of deep learning models in cross-scene generalization.

Main Methods:

• A two-phase framework combining meta-knowledge learning and unsupervised domain adaptation.
• A deep difference network (DDN) encodes scene-independent temporal change knowledge.
• Self-training domain adaptation with iterative evolution and pseudo-labeling.

Main Results:

• Significant improvements in background subtraction performance on the CDnet2014 dataset.
• Outperforms top unsupervised and supervised algorithms for unseen scenes by 9% and 3%, respectively.
• Achieves a favorable processing speed of 70 frames per second.

Conclusions:

• The proposed framework effectively generalizes to unseen scenes without annotations.
• Meta-knowledge learning and domain adaptation are key to robust cross-scene background subtraction.
• This approach offers a practical solution for real-world video surveillance and monitoring applications.