Section: New Results

Enhancing the Accuracy of SLAM-based Localization Systems for Autonomous Driving

Participants : Zayed Alsayed, Anne Verroust-Blondet, Fawzi Nashashibi.

Computing a reliable and accurate pose for a vehicle in any situation is one of the challenges for Simultaneous Localization And Mapping methods (SLAM) methods. Based on the probabilistic form of the SLAM solution, SLAM methods suffer from systematic errors related to the linearization of the solution models. The accuracy of the SLAM method can be improved by estimating a correction to be applied to the SLAM output based on relevant information available from the SLAM algorithm. In [20] two approaches predicting corrections to be applied to SLAM estimations are proposed:

1) The first approach is designed for 2D SLAM methods, i.e. independently of the underlying SLAM process and sensor used, where we aim to reduce the errors due to the dynamical modeling during specific maneuvers.

2) The second method is designed to handle errors related to the probabilistic formulation of Maximum Likelihood SLAM approaches, and thus it is suitable for 2D Maximum Likelihood SLAM methods (i.e. no assumptions on the sensor used).

The validity of both approaches was proved through two experiments using different evaluation metrics and using different sensor characteristics.

More detail can be found in the thesis manuscript of Zayed Alsayed entitled "Characterizing the Robustness and Enhancing the Accuracy of SLAM-based Localization Systems for Autonomous Driving" (cf. [7]).