Confomalized Multiband Uncertainty Regression and Reasoning for Visual Localization

This thesis introduces a novel uncertainty estimator designed to forecast multiband uncertainty intervals in robot pose estimation. The methodology involves the combination of the conformal prediction technique with a deep-learning regressor tailored explicitly for visual localization tasks. In this context, visual localization refers to the process of determining the robot's pose, retrieving its position and orientation within a predefined space constructed from previously gathered visual data. This process relies on monocular RGB images captured by its onboard camera as input. One of the peculiarities of our work is the use of conformal prediction to extract uncertainty intervals of the pose predicted by the model. Conformal prediction is a statistical framework designed to provide reliable measures of uncertainty by generating prediction intervals that guarantee a predefined confidence level. In our application, these intervals represent the uncertainty associated with the robot's pose predictions. Subsequently, we utilize the uncertainty estimations and optical flow-based reasoning process to enhance the final pose estimation accuracy. Tests show that our framework consistently outperforms traditional deep learning methods in challenging scenarios, such as scenarios characterized by significant noise, limited training data, and constrained model size, leading to a reduction in prediction error by a factor of 2-3.