Detecting shadows and low-lying objects in indoor and outdoor scenes using homographies

Many computer vision applications apply background suppression techniques for the detection and segmentation of moving objects in a scene. While these algorithms tend to work well in controlled conditions they often fail when applied to unconstrained real-world environments. This paper describes a system that detects and removes erroneously segmented foreground regions that are close to a ground plane. These regions include shadows, changing background objects and other low-lying objects such as leaves and rubbish. The system uses a set-up of two or more cameras and requires no 3D reconstruction or depth analysis of the regions. Therefore, a strong camera calibration of the set-up is not necessary. A geometric constraint called a homography is exploited to determine if foreground points are on or above the ground plane. The system takes advantage of the fact that regions in images off the homography plane will not correspond after a homography transformation. Experimental results using real world scenes from a pedestrian tracking application illustrate the effectiveness of the proposed approach.