Nonstationary Background Removal Via Multiple Camera Collaboration

An algorithm for nonstationary background removal is proposed for calibrated camera networks. Unlike conventional single-view background subtraction methods relying solely on the appearance or models of the background or the foreground, multiple-view networks introduce additional spatial information which can be used for extracting and validating the foreground of interest. At the local processing stage of the proposed algorithm, each camera creates a pseudo-background based on persistence of pixels in the collected image frames over time. For each incoming frame a noisy silhouette is produced by subtracting from the pseudo-background. In the collaborative part of the algorithm, we adopt the 3D voxel construction technique based on the noisy silhouettes to obtain the spatial information of the objects, which allows each camera to distinguish the foreground objects from the background objects, and hence removing the silhouettes corresponding to the background objects. In addition, if a part of the silhouette is created because of the noise on the image plane, inconsistency with observations from other cameras will result in the removal of it during the voxel construction. By removing the moving and stationary background objects, the proposed technique yields the image frames with the desired foreground on them. These foreground images can then be used for other applications such as target tracking or gesture analysis. Experimental results are provided to show that the proposed algorithm can remove the unwanted nonstationary background, and report the desired foreground successfully.