Local Processing as a Cue for Decreasing 3-D Structure Computation

Abstract

Usually, the extraction of 3-D information from a sequence of images taken by a mobile robot requires a lot of computation because of the complexity of both the 3-D geometry and the robot’s environment. If some powerful hardware for computing the 3-D structure in real time exists, its complexity cannot undergo a large-scale duplication. In this chapter, the authors stress that the problem of determining in close real time the 3-D structure of a mobile robot environment can be achieved by local computations. Such an approach implies the computation of local maps surrounding some points of interest selected in the scene, instead of the direct determination of the global one, as far as a local tracking rather than a global matching algorithm. For illustrating our purpose, we show the preliminary results of the 3-D local map reconstruction by an active monocular vision system mounted on a mobile robot. Such a technique computes and updates, while the robot is moving, the local map surrounding some feature points selected in the scene. Although the experimentation has succeeded only on stored data, its real-time implementation can be considered thanks to both the simplicity of equations and the use of a token tracking board that allows the processing of 14 images per second. The algorithmic and hardware aspects of this board are briefly reported, and its robustness is demonstrated by a robotic experiment.