SVD-based Tensor-Completion Technique for Background Initialization.

Area of Science:

• Computer Vision
• Signal Processing
• Machine Learning

Background:

• Background initialization is crucial for video analysis, especially with dynamic foregrounds.
• Existing methods struggle with complex scenes and computational efficiency.

Purpose of the Study:

• To propose a novel background-initialization technique using singular-value decomposition (SVD).
• To model scene backgrounds efficiently using rank-one matrices derived from video tensors.

Main Methods:

• Utilizes spatiotemporal slices from video tensors for SVD.
• Analyzes singular vectors and values to identify optimal background components.
• Proposes the spatiotemporal slice-based SVD (SS-SVD) method.

Main Results:

• A rank-1 matrix from the first SVD components effectively models the background.
• SS-SVD demonstrates superior performance across 93 complex video sequences.
• Achieves better results than state-of-the-art tensor/matrix completion, statistical, search, and labeling methods.

Conclusions:

• SS-SVD provides an efficient and effective solution for background initialization.
• The method offers reduced computational time and fewer frame requirements.
• Validated across diverse and challenging video datasets.