Application of CNN-based Reliability Estimation Method for Vision-based Self-localization

Abstract

We apply an existing reliability estimation method to vision-based self-localization in a RoboCup environment. The existing method has been used for LiDAR-based self-localization to measure the reliability of the estimation. For the judgement, a convolutional neural network (CNN) rates the consistency between the measurements from a LiDAR and an occupancy grid map. We replace the LiDAR and the map with a monocular color camera and a line map on a soccer field respectively. Our implementation was evaluated on a simulated soccer field. In the evaluation, it output appropriate evaluations toward the errors of vision-based Monte Carlo localization.