Mobile Robot Localization Considering Uncertainty of Depth Regression From Camera Images

Abstract

This letter presents a mobile robot localization method that uses depth regression from camera images. In this work, we use convolutional neural networks to regress the depth from the camera images. However, the depth regression results contain uncertainty, which must be resolved to stably perform localization with the depth regression results. This letter proposes a novel probabilistic model that enables the handling of the uncertainty of the depth regression results while localizing the robot pose. By handling the uncertainty, inaccurate depth regression results can be ignored, and localization robustness can be increased. We compare the proposed method with two traditional methods used in particle-filter-based localization that do not handle depth regression uncertainty. Comparison experiments are performed using two types of datasets: a manually created dataset using only a visual inertial odometry sensor, and the KITTI odometry dataset. Results show that the proposed method can track the robot pose even though the depth regression results are inaccurate and can increase localization accuracy without increasing memory and computational costs.