Robust Orthonormal Subspace Learning (ROSL): A Flexible Solution for Challenging Video Surveillance Scenarios
Sandeep Kumar, Vikrant Shokeen, Amit Kumar Sharma, Ahmad Taher Azar, Nashwa Ahmad Kamal, Chakib Ben Njima · 2025
Effectively differentiating between background and foreground objects is a key difficulty in video surveillance. The complexity of this task is heightened by the possibility of abnormalities, such as moving objects in crowded situations, fluctuating degrees of illumination, and noise. The model must therefore be flexible to accept changes to the scene. Existing models like Robust Principal Component Analysis (RPCA) with a convex relaxation have handled this topic by viewing it as a decomposition into low-rank and sparse matrices. However, as the data matrices expand, new difficulties appear, particularly in cases involving video analysis. Therefore, to solve these we concentrate on Robust Orthonormal Subspace Learning (ROSL), a novel methodology that finds a balance between complexity and limits. Several experiments using artificially noisy data, real-world videos, and facial images with various lighting situations have shown how effective this model is. We, also evaluate ROSL’s performance against the RPCA and the Singular Value Decomposition (SVD), respectively, which shows competitiveness.