Suppression of Detection Ghosts in Homography Based Pedestrian Detection

Abstract

One popular approach for multi-camera detection of pedestrians or other objects of interest in surveillance scenes is to perform background subtraction and project the resulting foreground mask images to a common scene plane using homographies. As the complexity of the scene increases, it is unavoidable that so called "ghost" detections should occur. These are false positives, indicating the presence of an object of interest where no such object actually exists. This paper proposes an approach to predicting where these ghost detections will occur, and provides a mechanism for suppressing their appearance.