Panoramic Video Separation with Online Grassmannian Robust Subspace Estimation

In this work, we propose a new total variation (TV)-regularized robust principal component analysis (RPCA) algorithm for panoramic video data with incremental gradient descent on the Grassmannian. The resulting algorithm has performance competitive with state-of-the-art panoramic RPCA algorithms and can be computed frame-by-frame to separate foreground/background in video with a freely moving camera and heavy sparse noise. We show that our algorithm scales favorably in computation time and memory. Finally we compare foreground detection accuracy and computation time of our method versus several existing methods.