3D image analysis for pedestrian detection

Abstract

A method for solving the dense disparity stereo correspondence problem is presented in this paper. This technique is designed specifically for pedestrian detection type applications. A new Ground Control Points (GCPs) scheme is introduced, using groundplane homography information to determine regions in which good GCPs are likely to occur. The method also introduces a dynamic disparity limit constraint to further improve GCP selection and dense disparity generation. The technique is applied to a real world pedestrian detection scenario with a background modeling system based on disparity and edges.