We found a match
Your institution may have access to this item. Find your institution then sign in to continue.
- Title
Research on optical flow visual odometer combined deep learning and geometric attitude determination.
- Authors
LI Peng; LIU Qi; HE Chen-yu; MAN Chun-tao; LIU Zong-ming
- Abstract
Aiming at the poor robustness and scale drift of multi-view geometry vision odometer, an optical flow vision odometer with deep learning and geometric attitude determination is proposed. Firstly, the optical flow and monocular depth between two images were estimated by deep learning. The 2D-2D correspondence was obtained from the optical flow estimation, and the camera relative pose was received by solving the essential matrix. Then, the triangulated depth was got from the pixel correspondence, and the predicted depth and the triangulated depth were scale-aligned adaptively. The depth error between the converted prediction depth and the triangulation depth was added to the loss function to optimize the depth network model, and the robust visual odometer DOF-VO was deduced to overcome the scale inconsistency. Extensive simulation experiments on the KITTI dataset show that the algorithm has a significant improvement compared to ORB_SLAM2 and SC-SfMLearner in the general evaluation index, and has good performance in trajectory tracking, which verifies the validity of the algorithm.
- Subjects
OPTICAL flow; ODOMETERS; DEEP learning; POSE estimation (Computer vision); COST functions; ATTITUDE (Psychology); TRIANGULATION; MONOCULARS
- Publication
Electric Machines & Control / Dianji Yu Kongzhi Xuebao, 2020, Vol 24, Issue 12, p142
- ISSN
1007-449X
- Publication type
Article
- DOI
10.15938/j.emc.2020.12.017