On the use of likelihood fields to perform sonar scan matching localization

Scan matching algorithms have been extensively used in the last years to perform mobile robot localization. Although these algorithms require dense and accurate sets of readings with which to work, such as the ones provided by laser range finders, different studies have shown that scan matching localization is also possible with sonar sensors. Both sonar and laser scan matching algorithms are usually based on the ideas introduced in the ICP (Iterative Closest Point) approach. In this paper a different approach to scan matching, the Likelihood Field based approach, is presented. Three scan matching algorithms based on this concept, the non filtered sNDT (sonar Normal Distributions Transform), the filtered sNDT and the LF/SoG (Likelihood Field/Sum of Gaussians), are introduced and analyzed. These algorithms are experimentally evaluated and compared to previously existing ICP-based algorithms. The obtained results suggest that the Likelihood Field based approach compares favorably with algorithms from the ICP family in terms of robustness and accuracy. The convergence speed, as well as the time requirements, are also experimentally evaluated and discussed.