SEMINAIRE DU 1 octobre 2015 – 16H @ LTCI – Salle C49
Low-Rank Video Segmentation for Background Estimation, and Multi-Temporal Foreground Detection in Videos
Background and foreground separation in videos is a ubiquitous task in many computer vision problems such as object recognition to tracking, and consequently much work has been done on this subject.
Recently, Candès et al. showed that such an estimation task could be addressed as a convex optimisation problem. The framework which they proposed is called Robust Principle Component Analysis (RPCA), and separates a matrix into the sum of a low-rank matrix (the background) and a sparse matrix (the foreground). The low-rank nature of the first matrix means that global lighting changes are handled. In this talk, we adapt and improve upon this framework in two ways.
Firstly, we propose an algorithm to achieve a « local » version of RPCA. Indeed, a considerable drawback of the standard RPCA is its poor ability to handle local lighting changes. We propose to model the background as piece-wise low-rank, each « piece » corresponding to spatio-temporally localised lighting conditions. The main challenge in this task is to identify the coherent regions, which we do with a region-merging approach based on spectral (graph) clustering. We show that this local RPCA allows for greater robustness to both local lighting changes and foreground elements which may remain static for a certain time.
Secondly, we present an online version of RPCA for the task of detecting foreground at multiple timescales. The goal of this analysis is to identify objects which remain in a certain position for a (user-defined) timescale. With visual examples, we show that this algorithm can be used for applications such as detecting the fluidity of traffic or detecting immobile people who may require assistance.
Aladair Newson is currently a postdoctoral researcher at Université Paris-Descartes.