Despite the recent advancements in Autonomous Vehicle (AV) technology, safety still remains a key challenge for their commercialisation and development. One of the major systems influencing the safety of AVs is its navigation system. Road localisation of autonomous vehicles is reliant on consistent accurate Global Navigation Satellite System (GNSS) positioning information. The GNSS relies on a number of satellites to perform triangulation and may experience signal loss around tall buildings, bridges, tunnels, trees, etc. We previously proposed the Wheel Odometry Neural Network (WhONet) as an approach to provide continuous positioning information in the absence of the GNSS signals. We achieved this by integrating the GNSS output with the wheel encoders’ measurements from the vehicle whilst also learning the uncertainties present in the position estimation. However, the positioning problem is a safety critical one and thus requires a qualitative assessment of the reasons for the predictions of the WhONet model at any point of use. There is therefore the need to provide explanations for the WhONet’s predictions to justify its reliability and thus provide a higher level of transparency and accountability to relevant stakeholders. Explainability in this work is achieved through the use of Shapley Additive exPlanations (SHAP) to examine the decision-making process of the WhONet model on an Inertial and Odometry Vehicle Navigation Benchmark Data subset describing an approximate straight-line trajectory. Our study shows that on an approximate straight-line motion, the two rear wheels are responsible for the most increase in the position uncertainty estimation error compared to the two front wheels.
Explainable Machine Learning for Autonomous Vehicle Positioning Using SHAP
Intel.Syst.Ref.Library
20.10.2022
27 pages
Aufsatz/Kapitel (Buch)
Elektronische Ressource
Englisch
DataCite | 2023
|Performance - IC125$ Over Shap
British Library Online Contents | 1992
Transportation Research Record | 2023
|