Stable Monocular Visual Odometry Based on Optical Flow Matching

A monocular visual odometry method combining deep learning and geometry is proposed to address the sparsity and position discontinuity problems. An unsupervised optical flow estimation network is constructed to obtain high-quality dense optical flow. Next, optical flow masks and depth masks are proposed for filtering key points. Finally, position estimation and scale recovery are performed based on multi-view geometry. Experiments are extensively validated on the KITTI dataset, and the visual odometry method achieves 3.48 (%) translation error and 0.67 rotation error ( \(^{\circ }\) /100 m) outperforming the baseline DF-VO.