Hierarchical 2.5-D Scene Alignment for Change Detection With Large Viewpoint Differences

Abstract

Change detection from mobile platforms is a relevant topic in the field of intelligent vehicles and has many applications, such as countering improvised explosive devices (C-IED). Existing real-time C-IED systems are not robust against large viewpoint differences, which are unavoidable under realistic operating conditions in outdoor environments. To address this, we proposes a new hierarchical 2.5-D scene-alignment algorithm. First, the 3-D ground surface of the historic scene is reconstructed by polygons, onto which historic image-based texture is projected. By estimating the 3-D transformation between historic and live camera views, the historic scene can be rendered as if seen from the live camera viewpoint. To compensate for 3-D alignment and reconstruction imperfections, local pixel-accurate registration refinement is performed in 2-D. The proposed real-time 2.5-D method thereby combines the accuracy of a 2-D local image registration with the robustness of 3-D scene alignment. It was found that the resulting change detection system detects small changes of only $18 \times 18 \times 9 \text{ cm}$ at distances of 60 meters under large trajectory deviations of up to 2.5 meters.