Graduate Student Seminar: Background Modelling for Video Surveillance Data using Robust Matrix Factorization Techniques

In the modern era of big data, high-dimensional datasets routinely suffer from heavy contamination and gross outliers. This poses a critical hurdle for computer vision tasks like automated video surveillance, where cleanly separating static backgrounds from dynamic foregrounds is vital for real-time target tracking and anomaly detection. While low-rank matrix decomposition via classical Principal Component Analysis (PCA) and Singular Value Decomposition (SVD) offers a solution, they do not work under erratic lighting shifts, or camera tampering. To overcome this, we propose a robust matrix factorization framework based on the popular minimum divergence framework, which allows balancing between efficiency and robustness. We also talk about desirable theoretical properties like consistency, equivariance, normality, and breakdown behaviour. The formulation extends far beyond video analytics: it can be used for low-rank matrix completion, and collaborative filtering, which we demonstrate through latent factor recovery on a Netflix movie recommendation problem.