Using terrain data and dynamical models is a promising approach to map-based passenger vehicle localization. In this approach, dynamical models are extracted from terrain data collected by a vehicle with a known location. The dynamical models are stored as a “map” of the data onto other vehicles. These vehicles can then discern their own location by comparing the newly acquired terrain data against the pre-extracted models. This approach has been shown to be an effective method of localization. However, system noise remains a significant challenge, affecting both model extraction and localization.
This paper introduces a novel approach to model extraction that maximizes the robustness of the extracted model map to inertial measurement unit noise. Three mechanisms are employed. First, the model map is represented as a tiered tree, with models describing successively finer data decimations in lower tree levels. Second, during the extraction process, the models and the transitions between models are chosen to accentuate the outlier end point that denotes the transition event. Finally, the extracted models are forced to have specific properties that address the noise added by the inertial measurement unit.
An additional benefit of the presented algorithm is that it generates model maps independently given a fixed model order. This provides a convenient method of efficiently adding new information to the vehicle's map. The approach is tested using vehicle pitch data collected in State College, Pennsylvania, USA.