Monocular Visual Odometry Using Unsupervised Deep Learning

This paper proposes a new monocular visual odometry system based on unsupervised deep learning that can simultaneously estimate the relative 6-DoF (degree of freedom) poses and scene depth of monocular cameras from sequences of images. It jointly trains depth maps and relative camera poses without ground truth. The depth model uses the encoder-decoder architecture and the pose model uses RCNN (Recurrent Convolutional Neural Network) with attention mechanism innovatively. The system is trained by using stereo images and tested by using monocular images. Experiments show that this unsupervised monocular visual odometry system doesn't suffer from the scale ambiguity of monocular visual odometry and it outperforms some state of the art approaches.