Reconstructive and Discriminative Subspace Learning Approach for Bakground/Foreground Separation

Introduction

Background subtraction is one of the basic low- level operations in video analysis. The aim is to separate static information called "background" from the moving objects called "foreground". The background needs to be modeled and updated over time to allow robust foreground detection. Reconstructive subspace learning such as Principal Component Analysis (PCA) has been widely used in background modeling by significantly reducing the data’s dimension. However, reconstructive representations strive to be as informative as possible in terms of well approximating the original data. On the other hand, discriminative meth- ods such as Linear Discriminant Analysis (LDA) provides a supervised reconstruction of the data which will often give better classification results when compared to the reconstructive methods. In this paper, we offer to use and validate the combination of a reconstructive method with a discriminative one to model robustly the background. The objective is firstly to enable a robust model of the background and secondly a robust classification of pixels as background or foreground. Results on different datasets demonstrate the performance of the proposed approach.

Principle

The proposed method used a reconstructive method with a discriminative one combining the advantages of both. This is done by embedding the LDA classification into the PCA framework. The combined subspace consists in a truncated PCA subspace and a few additional vectors that include the discriminative information which would be lost by the discarted principal vectors. The subspace contains both reconstructive information in order to enable incremental learning and discriminative information to enable efficient classification. PCA is a well known reconstructive method which includes the reconstructive information that can approximate well the training data. LDA, on the other hand, is a discriminative method which only keeps the discriminative information about the images. While the LDA is recognized to be better suitable to PCA in recognition tasks, it is less suitable for incremental learning. Therefore both methods will be mixed in order to achieve the best results.

Publication

C. Marghes, T. Bouwmans, R. Vasiu, “Background Modeling and Foreground Detection via a Reconstructive and Discriminative Subspace Learning Approach”, International Conference on Image Processing, Computer Vision, and Pattern Recognition, IPCV 2012, Las Vegas, USA, July 2012. [pdf]