Confidence Tuned Localization through Learning in the Loop

Abstract:
Learning based methods for localization in robotics and automated vehicles are a topic of ongoing research. While the methods that are currently in use work with a variety of sensor setups and show remarkable performance in pose estimation, research into uncertainty estimation of learning based methods is limited. This work presents a flexible method that allows to take the confidence of the localization method into account during training. Leveraging uncertainty estimates derived from learned features, we reduce overconfidence of the model and improve pose tracking performance purely derived from the training data.