Self-Supervised Learning of Depth Maps for Autonomous Cars

In autonomous driving, the knowledge of scene depth is essential for the control of car navigation. The depth can be measured by dedicated sensors like sonar, radar, LiDAR, ToF camera or obtained by computation from videos collected by an RGB camera. In this paper, we propose a novel solution for estimating depth maps from videos by using a neural network with self-supervised learning. The novelty consists in using ground truth position measured by IMU sensors on-board instead of motion estimated pose. The fusion of data from IMU sensors and RGB camera results in a lighter network and shorter training times. Results on depth accuracy and odometry are given after tests on KITTI dataset.