Examenspresentation av Vamsi Krishna och Vineet Kothari

Examiner: Lars Hammarstrand

Abstract:
For safe and efficient autonomous navigation of trucks, accurately estimating the lateral velocity along with quantifying the uncertainty of the estimation is imperative. This thesis focuses on developing deep learning models, including LSTM and Transformer networks, coupled with uncertainty quantification techniques such as Ensembles and Bayesian approximation to accurately estimate the lateral velocity of trucks and trailers. Moreover, the research addresses various aspects, including comparing deep learning-based lateral velocity estimation with conventional approaches like Kalman filters, evaluating different methods for quantifying uncertainty, and assessing how well the simulated data generalize to real truck-trailer combinations. To estimate the lateral velocity, the transformer model turned out to be the best-performing model, surpassing the conventional model-based approach under certain conditions. The Monte Carlo Dropout method showed slightly better performance for uncertainty quantification compared to the Gaussian Ensemble. However, challenges in performance variability and reliability still persist. The effectiveness of the models decreased when trained on one road surface and tested on another, highlighting the importance of diversity in training data. The model trained on simulated data performed well in the case of familiar real data, emphasizing the need for careful design and generation of simulated data for this purpose. Overall, deep learning shows promise in state estimation applications in the automotive industry but requires further development and consideration for safety-critical applications.

Welcome!

Vamsi, Vineet and Lars