Voting spaces cooperation for 3D plane detection from monocular image sequences

This paper deals with 3D scene reconstruction from an on-board moving camera in the context of automatic driver assistance systems. The aim of our study is to detect any kind of parameterized surface from a moving camera without camera calibration or any prior knowledge about the vehicle egomotion. We assume that the 3D scene is a set of 3D planes that are classified into three categories according to their orientation: lateral planes (buildings), horizontal planes (the road) and frontal planes (moving cars or crossing pedestrians). We propose an iterative voting process that takes advantages of some specific iso-velocity curves properties in order to build a set of appropriate voting spaces. Each of them facilitates the detection of a specific plane model. A tough problem as the detection of a parameterized surface from a moving camera is reduced to an easy maxima finding in several voting spaces. We focus in this paper on the iterative scheme that allows to deal with several spaces at the same time. We choose to adapt an histogram splitting approach in order to achieve a complete plane detection process.