Unsupervised Deep Persistent Monocular Visual Odometry and Depth Estimation in Extreme Environments

In recent years, unsupervised deep learning ap-proaches have received a significant attention to estimate depthand visual odometry (VO) from unlabelled monocular imagesequences. However, their performance is limited in challengingenvironments due to perceptual degradation, occlusions andrapid motions. Moreover, the existing unsupervised methodssuffer from the lack of scale-consistency constraints acrossframes, which causes that the VO estimators fail to providepersistent trajectories over long sequences. In this study, wepropose a unsupervised monocular deep VO framework thatpredicts 6 degrees-of-freedom pose camera motion and depthmap of the scene from unlabelled RGB image sequences.We provide detailed quantitative and qualitative evaluationsof the proposed framework on a) a challenging dataset col-lected during the DARPA Subterranean challenge1; and b)the benchmark KITTI and Cityscapes datasets. The proposedapproach outperforms both traditional and state-of-the-artunsupervised deep VO methods providing better results for bothpose estimation and depth recovery. The presented approach ispart of the solution used by the COSTAR team participatingat the DARPA Subterranean Challenge