A Generic Feature Selection Method for Background Subtraction Using Global Foreground Models

Abstract

Abstract. Over the last few years, a wide variety of background subtraction algorithms have been proposed for the detection of moving objects in videos acquired with a static camera. While much effort have been devoted to the development of robust background models, the automatic spatial selection of useful features for representing the background has been neglected. In this paper, we propose a generic and tractable feature selection method. Interesting contributions of this work are the proposal of a selection process coherent with the segmentation process and the exploitation of global foreground models in the selection strategy. Experiments conducted on the ViBe algorithm show that our feature selection technique improves the segmentation results.

On the Role and the Importance of Features for Background Modeling and Foreground Detection

“…According to the literature in this area, feature selection has been less investigated in backgroud modeling and foreground detection methods with only 9 papers. Practically, only five approaches have so far been used in the literature: (1) Adaboost [138] used with the classifier-based background model [158][160] [159][296], (2) Realboost [433] used with the KDE model [392], (3) dynamic feature selection [237] with OR-PCA model [238], (4) generic feature selection [56] with the ViBe model [31], and 5) One-class SVM [454]. These different approaches and their characteristics are analyzed in Section 17.…”

“…The computational time is shortened significantly by making the process of training and classification simple. As pointed out by Braham et al [56], the drawbacks of these approaches are unrealistic assumptions about the statistical distributions of foreground features that are used, i.e. a uniform distribution is assumed for the gray value of foreground objects and serves as a basis for computing other feature distributions.…”

“…Experimental results [392] show that this approach outperforms the original KDE [130] by a slight margin. While this framework can be used with an arbitrary number of candidate features, it has a serious drawback as pointed out by Braham et al [56]. Indeed, the synthetic foreground examples used for boosting are generated randomly from a uniform distribution.…”

“…Indeed, all candidate features are assumed to be uniformly distributed for foreground objects. This assumption is not valid because of the wide variety of foreground statistical distributions among different features [56]. For example, gradient has a foreground probability density func-tion concentrated around low values while RGB color components have a wider foreground probability density function.…”

“…-Feature selection via a generic method: Like in Parag et al [392], Braham et al [56] proposed that the feature selection process only occurs during the training phase, to avoid extra computations during normal background subtraction operations. Thus, the training phase is divided into three parts: 1) The first part consists to accumulate images free of foreground objects which are further processed to build local background statistical models, 2) The second part uses an other sequence of images, which include foreground objects and use to build a global foreground statistical model for each candidate feature of the feature set, and 3) the third part selects the best feature/threshold combination which is a local process, meaning that it is performed for each pixel individually.…”