Background Subtraction Using Adaptive Singular Value Decomposition

Abstract

Under the direction of (Dr. Mohammad Tabanjeh, Dr. Sergio Da Silva, and Dr. Sanwar Ahmed) We describe an efficient algorithm for background subtraction in video frames using an adaptive singular value decomposition method. By maintaining a subspace model of the static background scene, each incoming frame can be analyzed for new information representing potential foreground objects. The background subspace is iteratively updated in a computationally efficient manner through selective updating of singular vectors and values. Additional, optimizations such as thresholding and periodic forced updates further improve adaptation performance. Both qualitative and quantitative analysis on test imagery demonstrate accurate extraction of moving foreground elements from the background with minimal false positives. The overall approach provides an accurate and robust algorithm for real-time automated background subtraction applicable to tasks such as video surveillance.