Discussion and Conclusion

This paper has presented a solution to the SLAM problem considering six degrees of freedom and creating 3D maps of outdoor environments. It is based on ICP scan matching, initial pose estimation using a coarse-to-fine strategy with an octree representation and closing loop detection. Using an aerial photo as ground truth, the 3D map shows very good correspondence with the mapped environment, which was confirmed by a ratio comparison between map features and the respective photo features.

Compared with related approaches from the literature [6,10,26,27,28,29] we do not use a feature representation of the environment. Furthermore our algorithm manages registration without fixed data association. In the data association step, SLAM algorithms decide which features correspond. Wrong correspondences result in unprecise or even inconsistent models. The global scan matching based relaxation computes corresponding points, i.e., closest points, in every iteration. Furthermore, we avoid using probabilistic representations to keep the computation time at a minimum. The model optimization is solved in a closed form, i.e., by direct pose transformation. As a result of these efforts, registration and closed loop detection of 77 scans each with ca. 100000 points took only about 10 minutes.

Fig. 8 compares the probabilistic SLAM approaches with ours on an abstract level as presented by Folkesson and Christensen [9]. Robot poses are labeled with $X_i$ whereas the landmarks are the $Y_i$. Lines with black dots correspond to adjustable connections, e.g., springs, which can be relaxed by the algorithms. In our system, the measurements are fixed and data association is repeatedly done using nearest neighbor search.

Figure 8: Abstract comparison of SLAM approaches. Left: Probabilistic methods. The robot poses $X_i$ as well as the positions of the associated landmarks $Y_i$ are given in terms of a probability distribution. Global optimization tries to relax the model, where the landmarks are fixed. Small black dots on lines mark adjustable distances. Right: Our method with absolute measurements $Y_i$ (note there are no black dots between scan poses and scanned landmarks). The poses $X_i$ are adjusted based on scan matching aiming at collapsing the landmark copies $Y_{i_k}$ for all landmarks $Y_i$. Data association is the search for closest points.

Needless to say, a lot of work remains to be done. We plan to further improve the computation time and to use sensor uncertainty models. In addition, semantic labels for sub-structures of the resulting point model will be extracted.