Real-Time Map Building from a Stereo Camera under Unconstrained 3D Motion

Abstract

In this paper, a method is proposed for building occupancy grid based maps from sequences of stereo images in real time. The stereo camera can move and rotate unconstrained in three dimensions. This distinguishes the method from all other mapping approaches, which constrain the sensor to move in two dimensions. This freedom of motion allows to perform mapping in many new application areas. Additionally, three dimensional information can be maintained in layers or fused into a two dimensional overview of the environment. This transition is based on changeable rules rather than constraints. The method uses fuzzy logic and eliminates many problems that arise from traditional, probabilistic approaches for occupancy grids. It is shown that the proposed method performs well, is memory efficient and suitable for incremental use in real-time. The map will be used to assist navigation of a tele-operated mobile robot and real-time virtual walkthroughs, which are generated from the same stereo and motion data.