Introduction

General task: a mobile robot starts at an unknown position in an unknown environment and incrementally builds a 3D model of the visited area.




Finding an effective way to accomplish this task is deeply related to some design choices:



1. robot sensors
   
   
2. environment representation.




Previous approaches

• Simultaneous Localization and Mapping (SLAM) is addressed by a mathematical formalization of the dynamic system formed by the robot location and the location of all the features in the map and by computing the status of these system (for example by using an Extended Kalman Filter).
  
  The need of considering all the relationships between the features in the map makes the use of these approaches computationally intractable in many real applications.
  
  
  
• Camera Motion Estimation is addressed by using a mathematical formalization of the camera device and solving a over-constrained linear system with features extracted from images (e.g. lines and corners from images, etc.).
  
  Motion estimation from two or three images at a time is affected by error accumulation over long sequences.




Common Problem:



point/feature correspondences (i.e. data association).




New approach

We have investigated a new approach in which we focus our attention at the solution of both the motion estimation and the data association problems at the same time.

The general idea is to divide the process in two steps:



1. rough motion prediction by odometry and stereo
   
   
2. fine data matching and motion estimation by a local search in the robot's pose space.





Applications

1. Environment 3D modelling by using Hough Localization and Image Correlation,
   under the assumption of planar motion and planar surfaces.
   
   
2. Camera motion estimation and 3D object modelling by using Large Baseline Stereo.