Master Thesis - Deep Learning based TOA Error Regression with Uncertainty Estimation

Background. Time of arrival (TOA) based radio systems are widely used for indoor localization, achieving accuracies in the centimeter range. However, their positioning performance is restricted due to bandwidth limitations and challenging propagation conditions. Complex environments like production lines can deteriorate the positioning performance dramatically: Absorptions, reflections, and scattering cause wrong TOA estimations, which lead to localization errors. Therefore neural network based approaches have been investigated to mitigate the TOA estimation errors considering environment-specific properties. Often, the channel impulse response (CIR) is used, as the CIR contains rich spatial information which help localization algorithms to improve the positioning accuracy e.g. by none-line-of-sight (NLOS) identification, error mitigation, or TOA estimation.

Challenges. However, data-driven methods often have a poor generalization towards data-distribution shifts (in our case changed or new environments), which cause model violations and thus wrong predictions. Also, ambiguities in the data, caused, e.g., due to bandwidth limitations and NLOS cause a high uncertainty, which is often not considered in neural networks. To enable a reliable and robust localization performance, the uncertainty estimation of data-driven algorithms is therefore an important step.

Goals. The main goal of the work is to implement a TOA error mitigation approach employing channel impulse responses of radio systems. You will investigate and implement state of the art uncertainty estimation approaches like Monte Carlo dropout, deep ensembles or Bayesian neural networks to evaluate the reliability of the estimated errors.