Deep monocular visual odometry with generative adversarial network

Abstract

Recently, many technologies for autonomous driving are being developed. In particular, for the safety of a vehicle, a pose estimation method using various sensors is being developed. Among them, many studies using camera image data are being conducted. Images are data that can be obtained at a lower price than other sensors, and a lot of information can be utilized depending on the data processing method. The purpose of this thesis is to learn LSGAN, a kind of productive adversarial neural network, in a semi-supervised learning method, and to achieve a visual driving measurement method using a monocular camera. The method using a monocular camera is more versatile than the method using a binocular camera because it can be easily installed in small robots or drones. In order to estimate the accurate pose of a monocular camera in the conventional method, prior information such as ground truth must be provided. On the other hand, using deep learning, accurate pose estimation is possible without ground truth data of the situation.