FSO-VO: Visual odometry based on dense optical flow prediction and sequence optimization

The accuracy of geometry-based visual odometry methods is contingent upon high-quality feature points and scene-specific designs. When rapid motion significantly reduces the overlapping field of view between frames, these methods can incur substantial pose estimation errors due to an insufficient number of feature points, whereas purely deep learning-based approaches generally struggle to ensure accuracy. In addition, the cumulative drift problem and the scale drift problem inherent in monocular systems have not been effectively solved. In recent years, deep learning methods have shown great potential in optical flow estimation and depth estimation. Therefore, this paper explores an effective integration of deep learning with traditional geometric methods and proposes a robust monocular VO algorithm, FSO-VO. Specifically, a new optical flow extraction network (CA-Flow) is designed, which uses coordinate attention combined with graph structure information to provide robust and accurate feature points for pose calculation, and solves the scale drift problem by combining the scale estimated by the deep network. Then, the reprojection error is combined with the smoothing error, and the local sequence information is used to optimize the current posture, effectively reducing the cumulative drift problem. Experimental results on the KITTI dataset show that our method achieves the optimal accuracy on most sequences and has the best robustness on all sequences.